Thursday Mar 27, 2014

Why choose Oracle for Advanced Analytics?

If you're an enterprise company, chances are you have your data in an Oracle database. You chose Oracle for it's global reputation at providing the best software products (and now engineered systems) to support your organization. Oracle database is known for stellar performance and scalability, and Oracle delivers world class support.

If your data is already in Oracle Database or moving in that direction, leverage the high performance computing environment of the database to analyze your data. Traditionally it was common practice to move data to separate analytic servers for the explicit purpose of model building. This is no longer necessary nor is it scalable as your organization seeks to deliver value from Big Data. Oracle database now has several state of the art algorithms that execute in a parallel and distributed architecture directly in-database and augmented by custom algorithms in the R statistical programming language. Leveraging Oracle database for Advanced Analytics has benefits including:


  • Eliminates data movement to analytic servers

  • Enables analysis of all data not just samples

  • Puts your database infrastructure to even greater use

  • Eliminates impedance mismatch in the form of model translation when operationalizing models

  • All aspects of modeling and deployment are optionally available via SQL making integration into other IT software

  • Leverage CRAN algorithms directly in the database

Customers such as Stubhub, dunnhumby, CERN OpenLab, Financiera Uno, Turkcell, and others leverage Oracle Advanced Analytics to scale their applications, simplify their analytics architecture, and reduce time to market of predictive models from weeks to hours or even minutes.

Oracle leverages its own advanced analytics products, for example, by using Oracle Advanced Analytics in a wide range of Oracle Applications and internal deployments, ranging from:


  • Human Capital Management with Predictive Workforce to produce employee turnover, performance prediction, and "what if" analysis

  • Customer Relationship Management with Sales Prediction Engine to predict sales opportunities, what to sell, how much, and when

  • Supply Chain Management with Spend Classification to flag non-compliance or anomalies in expense submissions

  • Retail Analytics with Oracle Retail Customer Analytics to perform shopping cart analysis and next best offers

  • Oracle Financial Services Analytic Applications to enable quantitative analysts in credit risk management divisions to author rules/models directly in R


Oracle wants you to be successful with advanced analytics. Working closely with customers to integrate Oracle Advanced Analytics as an integral process of their analytics strategy, customers are able to put their advanced analytics into production much faster.

Thursday Mar 20, 2014

ROracle 1-1.11 released - binaries for Windows and other platforms available on OTN


We are pleased to announce the latest update of the open source ROracle package, version 1-1.11, with enhancements and bug fixes. ROracle provides high performance and scalable interaction from R with Oracle Database. In addition to availability on CRAN, ROracle binaries specific to Windows and other platforms can be downloaded from the Oracle Technology Network. Users of ROracle, please take our brief survey. We want to hear from you!

Latest enhancements in version 1-1.11 of ROracle:

• Performance enhancements for RAW data types and large result sets
• Ability to cache the result set in memory to reduce memory consumption on successive reads
• Added session mode to connect as SYSDBA or using external authentication
• bug 17383542: Enhanced dbWritetable() & dbRemoveTable() to work on global schema

Users of ROracle are quite pleased with the performance and functionality:


"In my position as a quantitative researcher, I regularly analyze database data up to a gigabyte in size on client-side R engines. I switched to ROracle from RJDBC because the performance of ROracle is vastly superior, especially when writing large tables. I've also come to depend on ROracle for transactional support, pulling data to my R client, and general scalability. I have been very satisfied with the support from Oracle -- their response has been prompt, friendly and knowledgeable."

           -- Antonio Daggett, Quantitative Researcher in Finance Industry


"Having used ROracle for over a year now with our Oracle Database data, I've come to rely on ROracle for high performance read/write of large data sets (greater than 100 GB), and SQL execution with transactional support for building predictive models in R. We tried RODBC but found ROracle to be faster, much more stable, and scalable."

           -- Dr. Robert Musk, Senior Forest Biometrician, Forestry Tasmania


See the ROracle NEWS for the complete list of updates.

We encourage ROracle users to post questions and provide feedback on the Oracle R Technology Forum.

In addition to being a high performance database interface to Oracle Database from R for general use, ROracle supports database access for Oracle R Enterprise.

Monday Mar 17, 2014

Oracle R Enterprise Upgrade Steps

We've recently announced that Oracle R Enterprise 1.4 is available on all platforms. To upgrade Oracle R Enterprise to the latest version:

  1. *Install the version of R that is required for the new version of Oracle R Enterprise.
         See the Oracle R Enterprise supported platforms matrix for the latest requirements.
  2. Update Oracle R Enterprise Server on the database server by running the install.sh script
         and follow the prompts for the upgrade path.
  3. Update the Oracle R Enterprise Supporting packages on the database server.
  4. Update the Oracle R Enterprise Client and Supporting packages on the client.

For RAC/Exadata installations, upgrade items 1, 2 and 3 must be performed on all compute notes. 

*If you've changed the R installation directory between releases, manually update the location of the R_HOME directory in the Oracle R Enterprise configuration table.  The sys.rqconfigset procedure edits settings in a configuration table called sys.rq_config. Use of this function requires the sys privilege. You can view the contents of this table to verify various environment settings for Oracle R Enterprise. Among the settings stored in sys.rq_config is the R installation directory, or R_HOME. The following query shows sample values stored in sys.rq_config for a Linux server:

SQL> select * from sys.rq_config;
NAME        VALUE
------------------------------------------------------------------
R_HOME       /usr/lib64/R
MIN_VSIZE    32M
MAX_VSIZE    4G
R_LIBS_USER  /u01/app/oracle/product/12.0.1/dbhome_1/R/library
VERSION      1.4
MIN_NSIZE    2M
MAX_NSIZE    20M

7 rows selected.

To point to the correct R_HOME:

SQL > sys.rqconfigset('R_HOME', '<path to current R installation directory>')

All Oracle R Enterprise downloads are available on the Oracle Technology Network. Refer to the instructions in section 8.3 of the Oracle R Enterprise Installations Guide for detailed steps on upgrading Oracle R Enterprise, and don't hesitate to post questions to the Oracle R forum.



Monday Mar 10, 2014

Oracle R Distribution 3.0.1 Benchmarks

Oracle R Distribution, Oracle's distribution of Open Source R, improves performance by dynamically linking to optimized, multi-threaded BLAS libraries. Unlike open source R, Oracle R Distribution uses all available cores and processors when dynamically linked against optimized BLAS, resulting in increased performance. Thus, the more cores available to Oracle R Distribution, the higher performance for many operations.

How is this possible?  Standard R's internal BLAS library was created when multi-core machines were not widely used, so it is single-threaded, i.e., operates on a single core. However, the BLAS API in R allows linking to different, multi-threaded BLAS libraries that allow linear algebra computations to use all cores and therefore run much faster. Oracle R Distribution simplifies the linking process by loading the high performance math library after it's added to PATH or LD_LIBRARY_PATH, depending on the Operating System.  Then you are set to use optimized math libraries - the Intel Math Kernel Library (MKL)AMD Core Math Library (ACML), or Solaris Sun Performance Library on Solaris. 

The benchmarks in this section demonstrate the performance of Oracle R Distribution 3.0.1 with and without dynamically loaded MKL. The R-25 benchmark script developed by the R community consists of fifteen tests. They are split into three groups (matrix calculation, matrix functions and "programmation") with trimmed means for each group, and each test is run three times. For this comparison, we report the mean for the three test runs. The benchmarks show that using Oracle R Distribution with dynamically loaded MKL libraries on a 30-core machine is significantly faster than the single core time.


Oracle R Distribution 3.0.1 Benchmarks


This benchmark was executed on a 3-node cluster, with 24 cores at 3.07GHz per CPU and 47 GB RAM, using Linux 5.5.


In-Database Scalability and Parallelism with Oracle R Enterprise


Oracle R Enterprise, the set of big data analytics R packages provided by Oracle, provides scalable, parallel in-database data manipulation and algorithms for analyzing very large data sets. Oracle R Enterprise functions implement parallel, out-of-core algorithms that overcome R's limitations of being memory-bound and single-threaded by executing requested R calculations on data in Oracle Database, using the database itself as the computational engine. Oracle R Enterprise allows users to further leverage Oracle's engineered systems, like Exadata, Big Data Appliance, and Exalytics, for enterprise-wide analytics, as well as reporting tools like Oracle Business Intelligence Enterprise Edition dashboards and BI Publisher documents. The combination of Oracle Database and R delivers an enterprise-ready, integrated environment for advanced analytics.

At the time of this post, Oracle R Distribution 3.0.1 is certified with Oracle R Enterprise 1.4.  See the Install Guide's Oracle R Enterprise support matrix  for a list of current Oracle R Distribution supported configurations and platforms, and this link for instructions on enabling high performance library support for Oracle R Distribution on a Windows or Linux client.

Tuesday Feb 18, 2014

Low-Rank Matrix Factorization in Oracle R Advanced Analytics for Hadoop

This guest post from Arun Kumar, a graduate student in the Department of Computer Sciences at the University of Wisconsin-Madison, describes work done during his internship in the Oracle Advanced Analytics group.

Oracle R Advanced Analytics For Hadoop (ORAAH), a component of Oracle’s Big Data Connectors software suite is a collection of statistical and predictive techniques implemented on Hadoop infrastructure. In this post, we introduce and explain techniques for a popular machine learning task that has diverse applications ranging from predicting ratings in recommendation systems to feature extraction in text mining namely matrix completion and factorization. Training, scoring, and prediction phases for matrix completion and factorization are available in ORAAH. The models generated can also be transparently loaded into R for ad-hoc inspection. In this blog, post we describe implementation specifics of these two techniques available in ORAAH.

Motivation

Consider an e-commerce company that displays products to potential customers on its webpage and collects data about views, purchases, ratings (e.g., 1 to 5 stars), etc. Increasingly, such online retailers are using machine learning techniques to predict in advance which products a customer is likely to rate highly and recommend such products to the customers in the hope that they might purchase them. Users build a statistical model based on the past history of ratings by all customers on all products. One popular model to generate predictions from such a hyper-sparse matrix is the latent factor model, also known as the low-rank matrix factorization model (LMF).

The setup is the following – we are given a large dataset of past ratings (potentially in the billions), say, with the schema (Customer ID, Product ID, Rating). Here, Customer ID refers to a distinct customer, Product ID refers to a distinct product, and Rating refers to a rating value, e.g., 1 to 5. Conceptually, this dataset represents a large matrix D with m rows (number of customers) and n columns (number of products), where the entries are the available ratings. Notice that this matrix is likely to be extremely sparse, i.e., many ratings could be missing since most customers typically rate only a few products. Thus, the task here is matrix completion – we need to predict the missing ratings so that it can be used for downstream processing such as displaying the top recommendations for each customer.

The LMF model assumes that the ratings matrix can be approximately generated as a product of two factor matrices, L and R, which are much smaller than D (lower rank). The idea is that the product L * R will approximately reconstruct the existing ratings and also automatically predict the missing ratings in D. More precisely, for each available rating (i,j,v) in D, we have (L x R) [i,j] ≈ v, while for each missing rating (i',j') in D, the predicted rating is (L x R) [i',j']. The model has a parameter r, which dictates the rank of the factor matrices, i.e., L is m x r, while R is r x n.

Matrix Completion in ORAAH

LMF can be invoked out-of-the-box using the routine orch.lmf. An execution based on the above example is shown below. The dataset of ratings is in a CSV file on HDFS with the schema above (named “retail_ratings” here).


input <- hdfs.attach("retail_ratings")
fit <- orch.lmf(input)

# Export the model into R memory
lr <- orch.export.fit(fit)

# Compute the prediction for the point (100, 50)

# First column of lr$L contains the userid
userid <- lr$L[,1] == 100 # find row corresponding to user id 100
L <- lr$L[, 2:(rank+1)]

#First column contains the itemid
itemid <- lr$R[,1] == 50 # find row corresponding to item id 50
R <- lr$R[, 2:(rank+1)]

# dot product as sum of terms obtained through component wise multiplication
pred <- sum(L[userid,] * R[itemid,])

The factor matrices can be transparently loaded into R for further inspection and for ad-hoc predictions of specific customer ratings using R. The algorithm we use for training the LMF model is called Incremental Gradient Descent (IGD), which has been shown to be one of the fastest algorithms for this task [1, 2].

The entire set of arguments for the function orch.lmf along with a brief description of each and their default values is given in the table below. The latin parameter configures the degree of parallelism for executing IGD for LMF on Hadoop [2]. ORAAH sets this automatically based on the dimensions of the problem and the memory available to each Mapper. Each Mapper fits its partition of the model in memory, and the multiple partitions run in parallel to learn different parts of the model. The last five parameters configure IGD and need to be tuned by the user to a given dataset since they can impact the quality of the model obtained.

ORAAH also provides routines for predicting ratings as well as for evaluating the model (computing the error of the model on a given labeled dataset) on a large scale over HDFS-resident datasets. The routine for prediction of ratings is predict, and for evaluating is orch.evaluate. Use help(orch.lmf) for online documentation, and demo(orch_lmf_jellyfish) for a fully working example including model fit, evaluation, and prediction.

Other Matrix Factorization Tasks

While LMF is primarily used for matrix completion tasks, it can also be used for other matrix factorization tasks that arise in text mining, computer vision, and bio-informatics, e.g., dimension reduction and feature extraction. In these applications, the input data matrix need not necessarily be sparse. Although many zeros might be present, they are not treated as missing values. The goal here is simply to obtain a low-rank factorization D ≈ L x R as accurately as possible, i.e., the product L x R should recover all entries in D, including the zeros. Typically, such applications use a Non-Negative Matrix Factorization (NMF) approach due to non-negativity constraints on the factor matrix entries. However, many of these applications often do not need non-negativity in the factor matrices. Using NMF algorithms for such applications leads to poorer-quality solutions. Our implementation of matrix factorization for such NMF-style tasks can be invoked out-of-the-box in ORAAH using the routine orch.nmf, which has the same set of arguments as LMF.

Experimental Results & Comparison with Apache Mahout

We now present an empirical evaluation of the performance, quality, and scalability of the ORAAH LMF tool based on IGD and compare it to the most widely used off-the-shelf tool for LMF on Hadoop – an implementation of the ALS algorithm from Apache Mahout [3].

All our experiments are run on an Oracle Big Data Appliance Hadoop cluster with nine nodes, each with Intel Xeon X5675 12-core 3.07GHz processors, 48 GB RAM, and 20 TB disk. We use 256MB HDFS blocks and 10 reducers for MapReduce jobs.

We use two standard public datasets for recommendation tasks – MovieLens10M (referred to as MLens) and Netflix – for the performance and quality comparisons (insert URL). To study scalability aspects, we use several synthetic datasets of different sizes by changing the number of rows, number of columns, and/or number of ratings. The table below presents the data set statistics.


Results: Performance and Quality

We first present end-to-end overview of the performance and quality achieved by our implementation and Mahout on MLens and Netflix. The rank parameter was set at 50 (a typical choice for such tasks) and the other parameters for both tools were chosen using a grid search. The quality of the factor matrices was determined using the standard measure of root mean square error (RMSE) [2]. We use a 70%-15%-15% Wold holdout of the datasets, i.e., 70% for training, 15% for testing, and 15% for validation of generalization error. The training was performed until 0.1% convergence, i.e., until the fractional decrease in the training RMSE after every iteration reached 0.1%. The table below presents the results.

1. ORAAH LMF has a faster performance than Mahout LMF on the overall training runtime on both datasets – 1.8x faster on MLens and 2.3x faster on Netflix.
2. The per-iteration runtime of ORAAH LMF is much lower than that of Mahout LMF – between 4.4x and 5.4x.
3. Although ORAAH LMF runs more iterations than Mahout LMF, the huge difference in the per-iteration runtimes make the overall runtime smaller for ORAAH LMF.
4. The training quality (training RMSE) achieved is comparable across both tools on both datasets. Similarly, the generalization quality is also comparable. Thus, ORAAH LMF can offer state-of-the-art quality along with faster performance.

Results: Scalability

The ability to scale along all possible dimensions of the data is key to big data analytics. Both ORAAH LMF and Mahout LMF are able to scale to billions of ratings by parallelizing and distributing computations on Hadoop. But we now show that unlike Mahout LMF, ORAAH LMF is also able to scale to hundreds of millions of customers (m) and products (n), and also scales well with the rank results along these three dimensions – m, n, and r. parameter (r, which affects the size of the factor matrices). The figure below presents the scalability.

1. Figures (A) and (B) plot the results for the Syn-row and Syn-col datasets, respectively (r = 2). ORAAH LMF scales linearly with both number of rows (m) and number of columns (n), while Mahout LMF does not show up on either plot because it crashes at all these values of m. In fact, we verified that Mahout LMF does not scale beyond even m = 20 M! The situation is similar with n. This is because Mahout LMF assumes that the factor matrices L and R fit entirely in the memory of each Mapper. In contrast, ORAAH LMF uses a clever partitioning scheme on all matrices ([2]) and can thus scale seamlessly on all dataset dimensions.
2. Figure (C) shows the impact of the rank parameter r. ORAAH LMF scales linearly with r and the per-iteration runtime roughly doubles between r = 20 and r = 100. However, the per-iteration runtime of Mahout LMF varies quadratically with r, and in fact, increases by a factor of 40x between r = 20 and r = 100! Thus, ORAAH LMF is also able to scale better with r.
3. Finally, on the tera-scale dataset Syn-tera with 1 billion rows, 10 million columns, and 20 billion ratings, ORAAH LMF (for r = 2) finishes an iteration in just under 2 hours!

Acknowledgements

The matrix factorization features in ORAAH were implemented and benchmarked by Arun Kumar during his summer internship at Oracle under the guidance of Vaishnavi Sashikanth. He is pursuing his PhD in computer science from the University of Wisconsin-Madison. This work is the result of a collaboration between Oracle and the research group of Dr. Christopher Ré, who is now at Stanford University. Anand Srinivasan helped integrate these features into ORAAH.

References

[1] Towards a Unified Architecture for in-RDBMS Analytics. Xixuan Feng, Arun Kumar, Benjamin Recht, and Christopher Ré. ACM SIGMOD 2012.

[2] Parallel Stochastic Gradient Algorithms for Large-Scale Matrix Completion. Benjamin Recht and Christopher Ré. Mathematical Programming Computation 2013.

[3] Apache Mahout. http://mahout.apache.org/.

Thursday Feb 13, 2014

Monitoring progress of embedded R functions

When you run R functions in the database, especially functions involving multiple R engines in parallel, you can monitor their progress using the Oracle R Enterprise datastore as a central location for progress notifications, or any intermediate status or results. In the following example, based on ore.groupApply, we illustrate instrumenting a simple function that builds a linear model to predict flight arrival delay based on a few other variables.

In the function modelBuildWithStatus, the function verifies that there are rows for building the model after eliminating incomplete cases supplied in argument dat. If not empty, the function builds a model and reports “success”, otherwise, it reports “no data.” It’s likely that the user would like to use this model in some way or save it in a datastore for future use, but for this example, we just build the model and discard it, validating that a model can be built on the data.


modelBuildWithStatus <-
  function(dat) {
    dat <- dat[complete.cases(dat),]
    if (nrow(dat)>0L) {
      mod <- lm(ARRDELAY ~ DISTANCE + AIRTIME + DEPDELAY, dat);
      "success"
    } else
      "no_data"
    }

When we invoke this using ore.groupApply, the goal is to build one model per “unique carrier” or airline. Using an ORE 1.4 feature, we specify the degree of parallelism using the parallel argument, setting it to 2.


res <- ore.groupApply(ONTIME_S[, c("UNIQUECARRIER","DISTANCE", "ARRDELAY", "DEPDELAY", "AIRTIME")],
        ONTIME_S$UNIQUECARRIER,
        modelBuildWithStatus,
        parallel=2L)

res.local<-ore.pull(res)
res.local[unlist(res.local)=="no_data"]

The result tells us about the status of each execution. Below, we print the unique carries that had no data.


R> res.local<-ore.pull(res)
R> res.local[unlist(res.local)=="no_data"]
$EA
[1] "no_data"

$`ML(1)`
[1] "no_data"

$`PA(1)`
[1] "no_data"

$PI
[1] "no_data"

$PS
[1] "no_data"

To monitor the progress of each execution, we can identify the group of data being processed in each function invocation using the value from the UNIQUECARRIER column. For this particular data set, we use the first two characters of the carrier’s symbol appended to “group.” to form a unique object name for storing in the datastore identified by job.name. (If we don’t do this, the value will form an invalid object name.) Note that since the UNIQUECARRIER column contains uniform data, we need only the first value.

The general idea for monitoring progress is to save an object in the datastore named for each execution of the function on a group. We can then list the contents of the named datastore and compute a percentage complete, which is discussed later in this post. For the “success” case, we assign the value “SUCCESS” to the variable named by the string in nm that we created earlier. Using ore.save, this uniquely named object is stored in the datastore with the name in job.name. We use the append=TRUE flag to indicate that the various function executions will be sharing the same named datastore.
If there is no data left in dat, we assign “NO DATA” to the variable named in nm and save that. Notice in both cases, we’re still returning “success” or “no data” so these come back in the list returned by ore.groupApply. However, we can return other values instead, e.g., the model produced.


modelBuildWithMonitoring <-
  function(dat, job.name) {
  nm <- paste("group.", substr(as.character(dat$UNIQUECARRIER[1L]),1,2), sep="")
  dat <- dat[complete.cases(dat),]
  if (nrow(dat)>0L) {
    mod <- lm(ARRDELAY ~ DISTANCE + AIRTIME + DEPDELAY, dat);
    assign(nm, "SUCCESS")
    ore.save(list=nm, name=job.name, append=TRUE)
    "success"
  } else {
    assign(nm, "NO DATA")
    ore.save(list=nm, name=job.name, append=TRUE)
    "no data"
  }
}

When we use this function in ore.groupApply, we provide the job.name and ore.connect arguments as well. The variable ore.connect must be set to TRUE in order to use the datastore. As the ore.groupApply executes, the datastore named by job.name will be increasingly getting objects added with the name of the carrier. First, delete the datastore named “job1”, if it exists.


ore.delete(name="job1")

res <- ore.groupApply(ONTIME_S[, c("UNIQUECARRIER","DISTANCE", "ARRDELAY", "DEPDELAY", "AIRTIME")],
        ONTIME_S$UNIQUECARRIER,
        modelBuildWithMonitoring,
        job.name="job1", parallel=2L, ore.connect=TRUE)

To see the progress during execution, we can use the following function, which takes a job name and the cardinality of the INDEX column to determine the percent complete. This function is invoked in a separate R engine connected to the same schema. If the job name is found, we print the percent complete, otherwise stop with an error message.


check.progress <- function(job.name, total.groups) {
  if ( job.name %in% ore.datastore()$datastore.name )
    print(sprintf("%.1f%%", nrow(ore.datastoreSummary(name=job.name))/total.groups*100L))
  else
    stop(paste("Job", job.name, " does not exist"))
}

To invoke this, compute the total number of groups and provide this and the job name to the function check.progress.
total.groups <- length(unique(ONTIME_S$UNIQUECARRIER))
check.progress("job1",total.groups)

However, we really want a loop to report on the progress automatically. One simple approach is to set up a while loop with a sleep delay. When we reach 100%, stop. To be self-contained, we include a simplification of the function above as a local function.


check.progress.loop <- function(job.name, total.groups, sleep.time=2) {
  check.progress <- function(job.name, total.groups) {
    if ( job.name %in% ore.datastore()$datastore.name )
      print(sprintf("%.1f%%", nrow(ore.datastoreSummary(name=job.name))/total.groups*100L))
    else
      paste("Job", job.name, " does not exist")
  }
  while(1) {
    try(x <- check.progress(job.name,total.groups))
    Sys.sleep(sleep.time)
    if(x=="100.0%") break
  }
}

As before, this function is invoked in a separate R engine connected to the same schema.


check.progress.loop("job1",total.groups)

Looking at the results, we can see the progress reported at one second intervals. Since the models build quickly, it doesn’t take long to reach 100%. For functions that take longer to execute or where there are more groups to process, you may choose a longer sleep time. Following this, we look at the datastore “job1” using ore.datastore and its contents using ore.datastoreSummary.


R> check.progress.loop("job1",total.groups,sleep.time=1)
[1] "6.9%"
[1] "96.6%"
[1] "100.0%"

R> ore.datastore(name="job1")
  datastore.name object.count size      creation.date description
1 job1 29 1073 2014-02-13 22:03:20
R> ore.datastoreSummary(name="job1")
object.name class size length row.count col.count
1 group.9E character 37 1 NA NA
2 group.AA character 37 1 NA NA
3 group.AQ character 37 1 NA NA
4 group.AS character 37 1 NA NA
5 group.B6 character 37 1 NA NA
6 group.CO character 37 1 NA NA
7 group.DH character 37 1 NA NA
8 group.DL character 37 1 NA NA
9 group.EA character 37 1 NA NA
10 group.EV character 37 1 NA NA
11 group.F9 character 37 1 NA NA
12 group.FL character 37 1 NA NA
13 group.HA character 37 1 NA NA
14 group.HP character 37 1 NA NA
15 group.ML character 37 1 NA NA
16 group.MQ character 37 1 NA NA
17 group.NW character 37 1 NA NA
18 group.OH character 37 1 NA NA
19 group.OO character 37 1 NA NA
20 group.PA character 37 1 NA NA
21 group.PI character 37 1 NA NA
22 group.PS character 37 1 NA NA
23 group.TW character 37 1 NA NA
24 group.TZ character 37 1 NA NA
25 group.UA character 37 1 NA NA
26 group.US character 37 1 NA NA
27 group.WN character 37 1 NA NA
28 group.XE character 37 1 NA NA
29 group.YV character 37 1 NA NA

The same basic technique can be used to note progress in any long running or complex embedded R function, e.g., in ore.tableApply or ore.doEval. At various points in the function, sequence-named objects can be added to a datastore. Moreover, the contents of those objects can contain incremental or partial results, or even debug output.

While we’ve focused on the R API for embedded R execution, the same functions could be invoked using the SQL API. However, monitoring would still be done from an interactive R engine.

Tuesday Feb 04, 2014

Invoking R scripts via Oracle Database: Theme and Variation, Part 6

How can I use "group apply" to partition data over multiple columns for parallel execution?
How can I use R for statistical computations and return results as a database table?

In this blog post of our theme and variation series, we answer these two questions through several examples, highlighting both R and SQL interfaces.

So far in this blog series on Oracle R Enterprise embedded R execution we've covered:

Part 1: ore.doEval / rqEval
Part 2: ore.tableApply / rqTableEval
Part 3: ore.groupApply / “rqGroupApply”
Part 4: ore.rowApply / rqRowEval
Part 5: ore.indexApply

Using ore.groupApply for partitioning data on multiple columns

While the “group apply” functionality is quite powerful as it is, users sometimes want to partition data on multiple columns. Since ore.groupApply currently takes only a single column for the INDEX argument, users can create a new column that is the concatenation of the columns of interest, and provide this column to the INDEX argument. We’ll illustrate this first using the R API, and then the SQL API.

R API

We adapt an example from Part 3 to illustrate partitioning data on multiple columns. Instead of building a C5.0 model, we’ll use the same CHURN_TRAIN data set, but build an rpart model since it will produce rules on the partitions of data we’ve chosen for the example, namely, voice_mail_plan and international_plan. To understand the number of rows we can expect in each partition, we’ll use the R table function. We then add a new column that pastes together the two columns of interest to create a new column called “vmp_ip”.


library(C50)
data(churn)

ore.create(churnTrain, "CHURN_TRAIN")

table(CHURN_TRAIN$international_plan, CHURN_TRAIN$voice_mail_plan)
CT <- CHURN_TRAIN
CT$vmp_ip <- paste(CT$voice_mail_plan,CT$international_plan,sep="-")
head(CT)

Each invocation of the function “my.rpartFunction” will receive data from one of the partitions identified in vmp_ip. Since our source partition columns are constants, we set them to NULL. The character vectors are converted to factors and the model is built to predict churn and saved in an appropriately named datastore. Instead of returning TRUE as done in the previous example, we create a list to return the specific partition column values, the distribution of churn values, and the model itself.


ore.scriptDrop("my.rpartFunction")
ore.scriptCreate("my.rpartFunction",
  function(dat,datastorePrefix) {
    library(rpart)
    vmp <- dat[1,"voice_mail_plan"]
    ip <- dat[1,"international_plan"]
    datastoreName <- paste(datastorePrefix,vmp,ip,sep="_")
    dat$voice_mail_plan <- NULL
    dat$international_plan <- NULL
    dat$state <- as.factor(dat$state)
    dat$churn <- as.factor(dat$churn)
    dat$area_code <- as.factor(dat$area_code)
    mod <- rpart(churn ~ ., data = dat)
    ore.save(mod, name=datastoreName, overwrite=TRUE)
    list(voice_mail_plan=vmp,
        international_plan=ip,
        churn.table=table(dat$churn),
        rpart.model = mod)
  })

After loading the rpart library and setting the datastore prefix, we invoke ore.groupApply using the derived column vmp_ip as the input to argument INDEX. After building the models, we’ll look at the first entry in the list returned. Using ore.load, we can load the model for the case where the customer neither has the voice mail plan, nor the international plan.


library(rpart)

datastorePrefix="my.rpartModel"

res <- ore.groupApply( CT, INDEX=CT$vmp_ip,
      FUN.NAME="my.rpartFunction",
      datastorePrefix=datastorePrefix,
      ore.connect=TRUE)
res[[1]]
ore.load(name=paste(datastorePrefix,"no","no",sep="_"))
mod
SQL API

To invoke this from the SQL API, we use the same approach as covered in Part 3. While we could create the table CT from the ore.frame used above, instead the following illustrates creating the derived column in SQL and explicitly defining a VIEW.


CREATE OR REPLACE VIEW CT AS
  SELECT t.*, "voice_mail_plan" || '-' || "international_plan" as "vmp_ip"
  FROM CHURN_TRAIN t;

Next, we create a PL/SQL PACKAGE and FUNCTION for the invocation.


CREATE OR REPLACE PACKAGE churnPkg AS
  TYPE cur IS REF CURSOR RETURN CT%ROWTYPE;
END churnPkg;
/
CREATE OR REPLACE FUNCTION churnGroupEval(
  inp_cur churnPkg.cur,
  par_cur SYS_REFCURSOR,
  out_qry VARCHAR2,
  grp_col VARCHAR2,
  exp_txt CLOB)
RETURN SYS.AnyDataSet
PIPELINED PARALLEL_ENABLE (PARTITION inp_cur BY HASH ("vmp_ip"))
CLUSTER inp_cur BY ("vmp_ip")
USING rqGroupEvalImpl;
/

Then, we can invoke the R function by name in the SELECT statement as follows:


select *
from table(churnGroupEval(
  cursor(select * from CT),
  cursor(select 1 as "ore.connect",' my.rpartModel2' as "datastorePrefix" from dual),
  'XML', 'state', 'my.rpartFunction'));

As another variation on this theme, suppose that you didn’t want to include all the columns from the source data set. To achieve this, you could create a view and define the PACKAGE from the view. However, you could also define a record that contains the specific columns of interest. This is a standard PL/SQL specification that can be used in combination with “group apply”.


CREATE OR REPLACE PACKAGE churnPkg2 AS
  TYPE rec IS RECORD ("vmp_ip" varchar2(8),
    "churn" varchar2(4),
    "state" varchar2(4),
    "account_length" NUMBER(38));
  TYPE cur IS REF CURSOR RETURN rec;
END churnPkg2;
/

If you don’t want to or cannot create a view, this allows you to specify the exact columns required for model building. Reducing the number of columns on input can improve performance, since only required data will be passed to the server-side R engine. Notice that we could have used this above since we remove the columns for the source partition columns.

How to return results from R statistical functions as database table data

R provides a wide range of statistical and advanced analytics functions. While Oracle Database contains a wide range of statistical functional in SQL, R further extends this set. In this next topic, we illustrate how to return statistical results as a SQL table for use with other SQL queries or to feed SQL-based applications.

As our example, we’ll use the R principal components function princomp. Our goal is to return the loadings of the PCA model as a database table. For our data set, we’ll use the USArrests data set provided with R. We can view the results of princomp in the mod variable, which has class “princomp”. We then push this data to Oracle Database, getting an ore.frame object.


mod <- princomp(USArrests, cor = TRUE)
class(mod)
mod
dat <- ore.push(USArrests)

R> mod <- princomp(USArrests, cor = TRUE)
R> class(mod)
[1] "princomp"
R> mod
Call:
princomp(x = USArrests, cor = TRUE)

Standard deviations:
   Comp.1    Comp.2    Comp.3    Comp.4
1.5748783 0.9948694 0.5971291 0.4164494

4 variables and 50 observations.
R> dat <- ore.push(USArrests)

In the first case considered, we use ore.tableApply to return simply the princomp object. When we do this we’re getting back a serialized object of type ore.object, but the actual princomp object still resides in the database. We can pull this object from the database to get a local princomp object, but this type of result cannot be directly returned as a SQL table because we need an object of class data.frame (which we’ll address later).


res <- ore.tableApply(dat,
      function(dat) {
        princomp(dat, cor=TRUE)
      })
class(res)
res.local <- ore.pull(res)
class(res.local)
str(res.local)
res.local
res

In the following output, we see the result is an ore.object that we pull from the database to get a princomp object. We examine the structure of the object and focus on the loadings element. In the example, we print res.local and res. Since res is an ore.object, it automatically gets pulled to the client before printing it.


R> res <- ore.tableApply(dat,
+ function(dat) {
+ princomp(dat, cor=TRUE)
+ })
R> class(res)
[1] "ore.object"
attr(,"package")
[1] "OREembed"
R> res.local <- ore.pull(res)
R> class(res.local)
[1] "princomp"
R> str(res.local)
List of 7
$ sdev : Named num [1:4] 1.575 0.995 0.597 0.416
..- attr(*, "names")= chr [1:4] "Comp.1" "Comp.2" "Comp.3" "Comp.4"
$ loadings: loadings [1:4, 1:4] -0.536 -0.583 -0.278 -0.543 0.418 ...
..- attr(*, "dimnames")=List of 2
.. ..$ : chr [1:4] "Murder" "Assault" "UrbanPop" "Rape"
.. ..$ : chr [1:4] "Comp.1" "Comp.2" "Comp.3" "Comp.4"
$ center : Named num [1:4] 7.79 170.76 65.54 21.23
..- attr(*, "names")= chr [1:4] "Murder" "Assault" "UrbanPop" "Rape"
$ scale : Named num [1:4] 4.31 82.5 14.33 9.27
..- attr(*, "names")= chr [1:4] "Murder" "Assault" "UrbanPop" "Rape"
$ n.obs : int 50
$ scores : num [1:50, 1:4] -0.986 -1.95 -1.763 0.141 -2.524 ...
..- attr(*, "dimnames")=List of 2
.. ..$ : chr [1:50] "1" "2" "3" "4" ...
.. ..$ : chr [1:4] "Comp.1" "Comp.2" "Comp.3" "Comp.4"
$ call : language princomp(x = dat, cor = TRUE)
- attr(*, "class")= chr "princomp"
R> res.local
Call:
princomp(x = dat, cor = TRUE)

Standard deviations:
   Comp.1    Comp.2    Comp.3    Comp.4
1.5748783 0.9948694 0.5971291 0.4164494

4 variables and 50 observations.
R> res
Call:
princomp(x = dat, cor = TRUE)

Standard deviations:
   Comp.1    Comp.2    Comp.3    Comp.4
1.5748783 0.9948694 0.5971291 0.4164494

4 variables and 50 observations.

In this next case, we focus on the loadings component of the princomp object, which contains the matrix of variable loadings, that is a matrix whose columns contain the eigenvectors. This is of class "loadings"…still not a data.frame. To convert the loadings component to a data.frame, we determine the dimensions of the matrix and then construct a data.frame by accessing the cells of the loading object. To get the variables associated with each row, we assign to the column variables the row names of the loadings. Finally, we return the loadings data.frame.


res <- ore.tableApply(dat,
      function(dat) {
        mod <- princomp(dat, cor=TRUE)
        dd <- dim(mod$loadings)
        ldgs <- as.data.frame(mod$loadings[1:dd[1],1:dd[2]])
        ldgs$variables <- row.names(ldgs)
        ldgs
      })
class(res)
res

In the output below, notice that we still have an ore.object being returned, but it’s in the form of a data.frame.


R> res <- ore.tableApply(dat,
+ function(dat) {
+ mod <- princomp(dat, cor=TRUE)
+ dd <- dim(mod$loadings)
+ ldgs <- as.data.frame(mod$loadings[1:dd[1],1:dd[2]])
+ ldgs$variables <- row.names(ldgs)
+ ldgs
+ })
R> class(res)
[1] "ore.object"
attr(,"package")
[1] "OREembed"
R> res
        Comp.1    Comp.2     Comp.3     Comp.4 variables
Murder -0.5358995 0.4181809 -0.3412327 0.64922780 Murder
Assault -0.5831836 0.1879856 -0.2681484 -0.74340748 Assault
UrbanPop -0.2781909 -0.8728062 -0.3780158 0.13387773 UrbanPop
Rape -0.5434321 -0.1673186 0.8177779 0.08902432 Rape

We can address this last issue by specifying the FUN.VALUE argument to get an ore.frame result (left as an exercise to the reader). But our main goal is to enable returning the loadings from SQL as a database table. For that, we create the function in the R script repository and construct the appropriate SQL query. In preparation for the next example, we’ll create the table USARRESTS using the R data set.


ore.create(USArrests,table="USARRESTS")

Now, we’ll switch to SQL. We’re introducing the functions sys.rqScriptDrop and sys.rqScriptCreate, which are used within a BEGIN END PL/SQL block, to store the R function ‘princomp.loadings’.


begin
--sys.rqScriptDrop('princomp.loadings');
sys.rqScriptCreate('princomp.loadings',
      'function(dat) {
        mod <- princomp(dat, cor=TRUE)
        dd <- dim(mod$loadings)
        ldgs <- as.data.frame(mod$loadings[1:dd[1],1:dd[2]])
        ldgs$variables <- row.names(ldgs)
        ldgs
      }');
end;
/

The SELECT statement provides input data by selecting all data from USARRESTS. There are no arguments to pass, so the next parameter is NULL. The SELECT string describes the format of the result. Notice that the column names must match in name (including case) and type. The last parameter is the name of the function stored in the R script repository.


select *
from table(rqTableEval( cursor(select * from USARRESTS),NULL,
          'select 1 as "Comp.1", 1 as "Comp.2", 1 as "Comp.3", 1 as "Comp.4", cast(''a'' as varchar2(12)) "variables" from dual','princomp.loadings'));

SQL> select *
from table(rqTableEval( cursor(select * from USARRESTS),NULL,
          'select 1 as "Comp.1", 1 as "Comp.2", 1 as "Comp.3", 1 as "Comp.4", cast(''a'' as varchar2(12)) "variables" from dual','princomp.loadings'));
2 3
    Comp.1     Comp.2     Comp.3     Comp.4 variables
---------- ---------- ---------- ---------- ------------
-.53589947 .418180865 -.34123273 .649227804 Murder
-.58318363 .187985604 -.26814843 -.74340748 Assault
-.27819087 -.87280619 -.37801579 .133877731 UrbanPop
-.54343209 -.16731864 .817777908 .089024323 Rape

If you have interesting embedded R scenarios to share with the ORE community, please consider posting a comment.

Monday Jan 20, 2014

Invoking R scripts via Oracle Database: Theme and Variation, Part 5


In the first four parts of Invoking R scripts via Oracle Database: Theme and Variation, we introduced features of Oracle R Enterprise embedded R execution involving the functions ore.doEval / rqEval, ore.tableApply / rqTableEval, ore.groupApply / “rqGroupApply”, and ore.rowApply / rqRowEval. In this blog post, we cover ore.indexApply. Note that there is no corresponding rqIndexEval – more on that later. The “index apply” function is also one of the parallel-enabled embedded R execution functions. It supports task-parallel execution, where one or more R engines perform the same or different calculations, or task. A number, associated with the index of the execution, is provided to the function. Any required data is expected to be explicitly generated or loaded within the function.

This functionality is valuable in a variety of settings, e.g., simulations, for taking advantage of high-performance computing hardware like Exadata.

As for “group apply” and “row apply”, Oracle Database handles the management and control of potentially multiple R engines at the database server machine, with only the index passed to the function as the first argument. Oracle Database ensures that each R function execution completes, otherwise the ORE function returns an error. Output formats as supported by the other embedded R functions are possible for ore.indexApply, for example, returning an ore.list or combining data.frame data into an ore.frame.

The variation on embedded R execution for ore.indexApply involves passing as an argument the number of times the user-defined R function should be executed.

Let’s look at a simple example.

The following code specifies to execute the function five times in parallel.


res <- ore.indexApply(5,
      function(index) {
        paste("IndexApply:",index)
      },
    parallel=TRUE)
class(res)
res

Notice that the class of the result is an ore.list, and when we print res, we have 5 character vectors, each with the index that was passed to the user-defined function. As with other parallel embedded R functions, the number of concurrently executing R engines can be limited by specifying the degree of parallelism of the database. As we’ll see in ORE 1.4, the parallel argument can specify a preferred number of parallel R engines, as an upper bound.


> class(res)
[1] "ore.list"
attr(,"package")
[1] "OREbase"
> res
$`1`
[1] "IndexApply: 1"

$`2`
[1] "IndexApply: 2"

$`3`
[1] "IndexApply: 3"

$`4`
[1] "IndexApply: 4"

$`5`
[1] "IndexApply: 5"

Column-parallel use case

If we wanted to parallelize R’s summary function, we could compute the summary statistics on each column in parallel and combine them into a final result. The following example does exactly that. While we could generalize this example, we focus on the iris data set and computing summary statistics on the first four numeric columns. Since iris comes standard with R, there’s no need to load data from any other source, we simply access it. The first argument to ore.indexApply is 4, the number of columns we wish to summarize in parallel. The function takes one argument, index, which will be a value between 1 and 4, and will be used to select the column to summarize. We massage the result of summary into a data.frame and add the column name to the result. Note that the function returns a single row: the summary statistics for the column.


res <- NULL
res <- ore.indexApply(4,
      function(index) {
        ss <- summary(iris[,index])
        attr.names <- attr(ss,"names")
        stats <- data.frame(matrix(ss,1,length(ss)))
        names(stats) <- attr.names
        stats$col <- names(iris)[index]
        stats
      },
      parallel=TRUE)
res

The result comes back as an ore.list object:


> res
$`1`
Min. 1st Qu. Median Mean 3rd Qu. Max. col
1 4.3 5.1 5.8 5.843 6.4 7.9 Sepal.Length

$`2`
Min. 1st Qu. Median Mean 3rd Qu. Max. col
1 2 2.8 3 3.057 3.3 4.4 Sepal.Width

$`3`
Min. 1st Qu. Median Mean 3rd Qu. Max. col
1 1 1.6 4.35 3.758 5.1 6.9 Petal.Length

$`4`
Min. 1st Qu. Median Mean 3rd Qu. Max. col
1 0.1 0.3 1.3 1.199 1.8 2.5 Petal.Width

This is good, but it would be better if the result was returned as an ore.frame, especially since all the columns are the same. To enable this, we’ll do a slight variation on the result by specifying FUN.VALUE with the structure of the result defined.


res <- ore.indexApply(4,
      function(index) {
        ss <- summary(iris[,index])
        attr.names <- attr(ss,"names")
        stats <- data.frame(matrix(ss,1,length(ss)))
        names(stats) <- attr.names
        stats$col <- names(iris)[index]
        stats
      },
      FUN.VALUE=data.frame(Min.=numeric(0),
        "1st Qu."=numeric(0),
        Median=numeric(0),
        Mean=numeric(0),
        "3rd Qu."=numeric(0),
        Max.=numeric(0),
        col=character(0)),
      parallel=TRUE)
res

Now, the result comes back as an ore.frame.


> res
  Min. X1st.Qu. Median  Mean X3rd.Qu. Max.      col
1 0.1 0.3 1.30 1.199 1.8 2.5 Petal.Width
2 1.0 1.6 4.35 3.758 5.1 6.9 Petal.Length
3 4.3 5.1 5.80 5.843 6.4 7.9 Sepal.Length
4 2.0 2.8 3.00 3.057 3.3 4.4 Sepal.Width
Simulation use case

The ore.indexApply function can be used in simulations as well. In this next example we take multiple samples from a random normal distribution with the goal to compare the distribution of the summary statistics. For this, we build upon the example above. We provide parameters such as the sample size, mean and standard deviation of the random numbers, and the number of simulations we want to perform. Each one of these simulations will occur in a separate R engine, in parallel, up to the degree of parallelism allowed by the database.

We specify num.simulations as the first parameter to ore.indexApply. Inside the user-defined function, we pass the index and three arguments to the function. The function then sets the random seed based on the index. This allows each invocation to generate a different set of random numbers. Using rnorm, the function produces sample.size random normal values. We invoke summary on the vector of random numbers, and then prepare a data.frame result to be returned. We’re using the FUN.VALUE to get an ore.frame as the final result.


res <- NULL
sample.size = 1000
mean.val = 100
std.dev.val = 10
num.simulations = 1000

res <- ore.indexApply(num.simulations,
      function(index, sample.size=1000, mean=0, std.dev=1) {
        set.seed(index)
        x <- rnorm(sample.size, mean, std.dev)
        ss <- summary(x)
        attr.names <- attr(ss,"names")
        stats <- data.frame(matrix(ss,1,length(ss)))
        names(stats) <- attr.names
        stats$index <- index
        stats
      },
      FUN.VALUE=data.frame(Min.=numeric(0),
        "1st Qu."=numeric(0),
        Median=numeric(0),
        Mean=numeric(0),
        "3rd Qu."=numeric(0),
        Max.=numeric(0),
        index=numeric(0)),
      parallel=TRUE,
      sample.size=sample.size,
      mean=mean.val, std.dev=std.dev.val)
res
boxplot(ore.pull(res[,1:6]),
  main=sprintf("Boxplot of %d rnorm samples size %d, mean=%d, sd=%d",
        num.simulations, sample.size, mean.val, std.dev.val))

To get the distribution of samples, we invoke boxplot on the data.frame after pulling the result to the client.

Here are a couple of plots showing results for different parameters:


In both cases, we run 10,000 samples. The first graph uses a sample size of 10 and the second uses a sample size of 1000. From these results, it is clear that a larger sample size significantly reduces the variance in each of the summary statistics - confirming our Statistics 101 understanding.

Error reporting

As introduced above, Oracle Database ensures that each embedded R user-defined function execution completes, otherwise the ORE function returns an error. Of course, any side-effects of the user-defined function need to be manually cleaned up. Operations that produce files, create tables in the database, or result in completed database transactions through ROracle will remain intact. The ORE embedded R infrastructure will report errors as produced by the function as illustrated in the following example.

The code specifies to invoke 4 parallel R engines. If the index has value 3, attempt to load the non-existant package "abc123" (which produces an error), otherwise return the index value.


R> ore.indexApply(4,
+ function(index) {
+ if (index==3) {library(abc123)}
+ else {return(index)}
+ }
+ )
Error in .oci.GetQuery(conn, statement, data = data, prefetch = prefetch, :
ORA-12801: error signaled in parallel query server P000
ORA-20000: RQuery error
Error in library(abc123) : there is no package called 'abc123'
ORA-06512: at "RQSYS.RQGROUPEVALIMPL", line 121
ORA-06512: at "RQSYS.RQGROUPEVALIMPL", line 118

Notice that the first reported error is an ORE-12801: error signaled in parallel query server. Then the ORA-20000: RQuery error indicates the error as returned by the R engine. Also interesting to note is that the ORA-06512 errors reveal the underlying implementation of ore.indexApply "RQSYS.RQGROUPEVALIMPL". Which leads us to the next topic.

No rqIndexEval?

“Index apply” is really a variation of “group apply” where the INDEX column is a numeric vector that is pushed to the database. With n distinct numbers, one number is provided to each function as its index. As a result, there is no corresponding rqIndexEval in the SQL API. The user would have to create a similar package and function as was illustrated in the blog post on “group apply.”

Thursday Jan 09, 2014

Invoking R scripts via Oracle Database: Theme and Variation, Part 4

In the first three parts of Invoking R scripts via Oracle Database: Theme and Variation, we introduced features of Oracle R Enterprise embedded R execution involving the functions ore.doEval / rqEval, ore.tableApply / rqTableEval, and ore.groupApply / “rqGroupApply”. In this blog post, we’ll cover the next in our theme and variation series involving ore.rowApply and rqRowEval. The “row apply” function is also one of the parallel-enabled embedded R execution functions. It supports data-parallel execution, where one or more R engines perform the same R function, or task, on disjoint chunks of data. This functionality is essential to enable scalable model scoring/predictions on large data sets and for taking advantage of high-performance computing hardware like Exadata.

As for ore.groupApply, Oracle Database handles the management and control of potentially multiple R engines at the database server machine, automatically chunking and passing data to parallel executing R engines. Oracle Database ensures that R function executions for all chunks of rows complete, or the ORE function returns an error. The result from the execution of each user-defined embedded R function is gathered in an ore.list. This list remains in the database until the user requires the result. However, we’ll also show how data.frame results from each execution can be combined into a single ore.frame. This features works for return values of other embedded R functions as well.

The variation on embedded R execution for ore.rowApply involves passing not only an ore.frame to the function such that the first parameter of your embedded R function receives a data.frame, but also the number of rows that should be passed to each invocation of the user-defined R function. The last chunk, of course, may have fewer rows than specified.

Let’s look at an example. We’re going to use the C50 package to score churn data (i.e., predict which customers are likely to churn) using the C5.0 decision tree models we built in the previous blog post with ore.groupApply. (Well, almost. We need to rebuild the models to take into account the full data set levels.) The goal is to score the customers in parallel leveraging the power of a high performance computing platform, such as Exadata.


library(C50)
data(churn)

ore.create(churnTest, "CHURN_TEST")

myFunction <- function(dat, xlevels, datastorePrefix) {
  library(C50)
  state <- dat[1,"state"]
  datastoreName <- paste(datastorePrefix,state,sep="_")
  dat$state <- NULL
  for (j in names(xlevels))
    dat[[j]] <- factor(dat[[j]], levels = xlevels[[j]])
  ore.load(name=datastoreName)
  res <- data.frame(pred=predict(mod,dat, type="class"),
        actual=dat$churn,
        state=state)
  res
}

xlevels <- ore.getXlevels(~ ., CHURN_TEST[,-1])
scoreList <- ore.groupApply(
  CHURN_TEST,
  INDEX=CHURN_TEST$state,
  myFunction,
  datastorePrefix="myC5.0model3",xlevels=xlevels, ore.connect=TRUE)
score.MA <- ore.pull(scoreList$MA)
table(score.MA$actual, score.MA$pred)

A few points to highlight:

• Instead of computing the levels using the as.factor function inside the user-defined function, we’ll use ore.getXlevels, which returns the levels for each factor column. We don’t need this for the state column, so we exclude it (“-1”). In the previous post we noted that factor data is passed as character columns in the data.frame. Computing the levels first can ensure that all possible levels are provided during model building, even if there are no rows with some of the level values.
• When building models where some levels were missing (due to using as.factor on each partition of data), scoring can fail if the test data has unknown level values. For this reason, the models built in Part 3 need to be rebuilt using the approach above with ore.getXlevels. This is left as an exercise for the reader.
• Assign the function to the variable “myFunction” to facilitate reuse (see below).
• We construct the datastore name to be the same as when we were building the models, i.e., appending the state value to the datastore prefix separated by an ‘_’.
• The for loop iterates over the levels passed in as xlevels, creating a factor using the provided levels and assigning it back to the data.frame.
• Loading the datastore by name, we have access to the variable mod, which contains the model for the particular state.
• The result is constructed as a data.frame with the prediction and the actual values.
• Three arguments are passed: the datastore prefix, the levels that were pre-computed, and that we need to connect to the database because we’re using a datastore.
• The results are stored as a list of ore.frames. We can pull the scores for MA and compute a confusion matrix using table.

This is fine. However, we likely don’t want to have a list of separate ore.frames as the result. We’d prefer to have a single ore.frame with all the results. This can be accomplished using the FUN.VALUE argument. Whenever a data.frame is the result of the user-defined R function, and if the structure of that data.frame is the same across all invocations of the group apply or row apply, you can combine them into a single result by defining the structure as follows:

scores <- ore.groupApply(
  CHURN_TEST,
  INDEX=CHURN_TEST$state,
  myFunction,
  datastorePrefix="myC5.0model3",xlevels=xlevels, ore.connect=TRUE,
  FUN.VALUE=data.frame(pred=character(0),
        actual=character(0),
        state=character(0)));
head(scores)
scores.local <- ore.pull(scores)
table(scores.local[scores.local$state=="MA",c("actual","pred")])

scores.MA <- scores[scores$state=="MA",c("actual","pred")]
table(scores.MA$actual, scores.MA$pred)

A few important points to highlight:

• FUN.VALUE is set to a data.frame that describes the format of the result. By providing this argument, you will get back a single ore.frame, not an ore.list object.
• The group apply completes instantaneously because it is only defining the ore.frame, not actually performing the scoring. Not until the values are needed does the result get computed. We invoke head on the ore.frame in scores to highlight this.
• We can pull the scores to the client to invoke table as before, but subselecting for state MA. However, we can also do this computation in the database using the transparency layer. First, we filter the rows for MA in scores.MA, and then invoke table on the two columns. Note: ORE requires passing the two columns explicitly to the overloaded function table.
• To do this in parallel, add the argument parallel=TRUE to the ore.groupApply call.

Wait! What happened to ore.rowApply?

Above, we showed how to score with multiple models using ore.groupApply. But what if we had customers from a single state that we wanted to score in parallel? We can use ore.rowApply and rqRowEval to invoke a function on chunks of data (rows) at a time, from 1 to the total number of rows. (Note that values closer to the latter will have no benefit from parallelism, obviously.)


scores <- ore.rowApply(
  CHURN_TEST[CHURN_TEST$state=="MA",],
  myFunction,
  datastorePrefix="myC5.0model3",xlevels=xlevels,
  ore.connect=TRUE, parallel=TRUE,
  FUN.VALUE=data.frame(pred=character(0),
        actual=character(0),
        state=character(0)),
  rows=200)
scores
table(scores$actual, scores$pred)

A few points to highlight:

• Since we want to perform the scoring in parallel by state, we filter the rows for MA. This will ensure that all rows processed can use the same predictive model.
• We set the rows argument to 200. CHURN_TEST has 1667 rows, so this will result in nine executions of myFunction. The first eight receiving 200 rows each and the last receiving 67 rows.
• We also set parallel=TRUE above since we want the scoring performed in parallel.
• The invocation of ore.rowApply returns immediately. Not until we print scores do we incur the cost of executing the underlying query. However, also note that each time we access scores, for example in the following call to table, we incur the cost of executing the query. If the result will be used many times in subsequent operations, you may want to create a table with the result using ore.create.

In SQL, we can do the same, but we’ll need to store the function in the R script repository (perhaps called "myScoringFunction") and also store xlevels in a datastore (perhaps called "myXLevels"). While we can pass complex objects in the R interface to embedded R functions, we cannot do that in SQL. Instead, we must pass the name of a datastore. Since the xlevels are in a datastore, the user-defined R function needs to be modified to take this other datastore name and load that datastore to have access to xlevels. This set of changes is left to the reader as an exercise.


select * from table(rqRowEval(
  cursor(select /*+ parallel(t, 4) */ *
        from CHURN_TEST t
        where "state" = 'MA'),
  cursor(select 1 as "ore.connect",
        'myC5.0model3' as "datastorePrefix",
        'myXLevels' as "xlevelsDatastore"
        from dual),
  'select ''aaa'' "pred",''aaa'' "actual" , ''aa'' "state" from dual',
    200, 'myScoringFunction'));

A few points to highlight:

• The input cursor specifies a parallel hint on the input data cursor and filtering data for MA as well.
• Several arguments are being passed, including the new argument to our function myXLevels.
• The output form is specified in the SQL string. Care must be taken to ensure that the column names, ordering, and the length of character strings match the returned data.frame.

Map Reduce

The “row apply” functionality can be thought of in terms of the map-reduce paradigm where the mapper performs the scoring and outputs a data.frame value (no key required). There is no reducer, or the reducer is simply a pass-through.

Memory and performance considerations

Unlike with group apply, the rows argument in row apply ensures an upper bound on the number of rows (and hence memory requirement). The value of rows should be chosen to balance memory and parallel performance. The usual measures can be taken regarding setting memory limits on the R engine – as noted in Part 2.

There may be instances where setting rows = 1 makes sense. For example, if the computation per row is intensive (i.e., takes a long time), sending one row per R engine may be appropriate. Experiment with a range of values for rows to determine the best value for your particular scenario.

Sunday Jan 05, 2014

Invoking R scripts via Oracle Database: Theme and Variation, Part 3

In the first two parts of Invoking R scripts via Oracle Database: Theme and Variation, we introduced features of Oracle R Enterprise embedded R execution, focusing on the functions ore.doEval / rqEval and ore.tableApply / rqTableEval. In this blog post, we’ll cover the next in our theme and variation series involving ore.groupApply and the corresponding definitions required for SQL execution. The “group apply” function is one of the parallel-enabled embedded R execution functions. It supports data-parallel execution, where one or more R engines perform the same R function, or task, on different partitions of data. This functionality is essential to enable the building of potentially 10s or 100s of thousands of predictive models, e.g., one per customer, and for taking advantage of high-performance computing hardware like Exadata.

Oracle Database handles the management and control of potentially multiple R engines at the database server machine, automatically partitioning and passing data to parallel executing R engines. It ensures that all R function executions for all partitions complete, or the ORE function returns an error. The result from the execution of each user-defined embedded R function is gathered in an ore.list. This list remains in the database until the user requires the result.

The variation on embedded R execution for ore.groupApply involves passing not only an ore.frame to the function such that the first parameter of your embedded R function receives a data.frame, but also an INDEX argument that specifies the name of a column by which the rows will be partitioned for processing by a user-defined R function.

Let’s look at an example. We’re going to use the C50 package to build a C5.0 decision tree model on the churn data set from C50. The goal is to build one churn model on the data for each state.


library(C50)
data(churn)

ore.create(churnTrain, "CHURN_TRAIN")

modList <- ore.groupApply(
  CHURN_TRAIN,
  INDEX=CHURN_TRAIN$state,
    function(dat) {
      library(C50)
      dat$state <- NULL
      dat$churn <- as.factor(dat$churn)
      dat$area_code <- as.factor(dat$area_code)
      dat$international_plan <- as.factor(dat$international_plan)
      dat$voice_mail_plan <- as.factor(dat$voice_mail_plan)
      C5.0(churn ~ ., data = dat, rules = TRUE)
    });
mod.MA <- ore.pull(modList$MA)
summary(mod.MA)

A few points to highlight:
• As noted in Part 2 of this series, to use the CRAN package C50 on the client, we first load the library, and then the churn data set.
• Since the data is a data.frame, we’ll create a table in the database with this data. Notice that if you compare the results of str(churnTrain) with str(CHURN_TRAIN), you will see that the factor columns have been retained. This becomes relevant later.
• The function ore.groupApply will return a list of models stored as ore.object instances. The first argument is the ore.frame CHURN_TRAIN and the second argument indicates to partition the data on column state such that the user-defined function is invoked on each partition of the data.
• The next argument specifies the function, which could alternatively have been the function name if the FUN.NAME argument were used and the function saved explicitly in the R script repository. The function’s first argument (whatever its name) will receive one partition of data, e.g., all data associated with a single state.
• Regarding the user-defined function body, we explicitly load the package we’re using, C50, so the function body has access to it. Recall that this user-defined R function will execute at the database server in a separate R engine from the client.
• Since we don’t need to know which state we’re working with and we don’t want this included in the model, we delete the column from the data.frame.
• Although the ore.frame defined functions, when they are loaded to the user-defined embedded R function, factors appear as character vectors. As a result, we need to convert them back to factors explicitly.
• The model is built and returned from the function.
• The result from ore.groupApply is a list containing the results from the execution of the user-defined function on each partition of the data. In this case, it will be one C5.0 model per state.
• To view the model, we first use ore.pull to retrieve it from the database and then invoke summary on it. The class of mod.MA is “C5.0”.

SQL API

We can invoke the function through the SQL API by storing the function in the R script repository. Previously we showed doing this using the SQL API, however, we can also do this using the R API , but we’re going to modify the function to store the resulting models in an ORE datastore by state name:


ore.scriptCreate("myC5.0Function",
  function(dat,datastorePrefix) {
    library(C50)
    datastoreName <- paste(datastorePrefix,dat[1,"state"],sep="_")
    dat$state <- NULL
    dat$churn <- as.factor(dat$churn)
    dat$area_code <- as.factor(dat$area_code)
    dat$international_plan <- as.factor(dat$international_plan)
    dat$voice_mail_plan <- as.factor(dat$voice_mail_plan)
    mod <- C5.0(churn ~ ., data = dat, rules = TRUE)
    ore.save(mod, name=datastoreName)
    TRUE
  })

Just for comparison, we could invoke this from the R API as follows:


res <- ore.groupApply( CHURN_TRAIN, INDEX=CHURN_TRAIN$state,
          FUN.NAME="myC5.0Function",
          datastorePrefix="myC5.0model", ore.connect=TRUE)
res
res <- ore.pull(res)
all(as.logical(res) == TRUE)

Since we’re using a datastore, we need to connect to the database setting ore.connect to TRUE. We also pass the datastorePrefix. The result res is an ore.list of logical values. To test if all are TRUE, we first pull the result and use the R all function.

Back to the SQL API…Now that we can refer to the function in the SQL API, we invoke the function that places one model per datastore, each with the given prefix and state.


select *
from table(churnGroupEval(
  cursor(select * from CHURN_TRAIN),
  cursor(select 1 as "ore.connect",' myC5.0model2' as "datastorePrefix" from dual),
  'XML', 'state', 'myC5.0Function'));

There’s one thing missing, however. We don’t have the function churnGroupEval. There is no generic “rqGroupEval” in the API – we need to define our own table function that matches the data provided. Due to this and the parallel nature of the implementation, we need to create a PL/SQL FUNCTION and supporting PACKAGE:


CREATE OR REPLACE PACKAGE churnPkg AS
  TYPE cur IS REF CURSOR RETURN CHURN_TRAIN%ROWTYPE;
END churnPkg;
/
CREATE OR REPLACE FUNCTION churnGroupEval(
  inp_cur churnPkg.cur,
  par_cur SYS_REFCURSOR,
  out_qry VARCHAR2,
  grp_col VARCHAR2,
  exp_txt CLOB)
RETURN SYS.AnyDataSet
PIPELINED PARALLEL_ENABLE (PARTITION inp_cur BY HASH ("state"))
CLUSTER inp_cur BY ("state")
USING rqGroupEvalImpl;
/

The highlights in red indicate the specific parameters that need to be changed to create this function for any particular data set. There are other variants, but this will get you quite far.

To validate that our datastores were created, we invoke ore.datastore(). This returns the datastores present and we will see 51 such entries – one for each state and the District of Columbia.

Parallelism

Above, we mentioned that “group apply” supports data parallelism. By default, parallelism is turned off. To enable parallelism, the parameter to ore.groupApply needs to be set to TRUE.


ore.groupApply( CHURN_TRAIN, INDEX=CHURN_TRAIN$state,
          FUN.NAME="myC5.0Function",
          datastorePrefix="myC5.0model",
          ore.connect=TRUE,
          parallel=TRUE
)

In the case of the SQL API, the parallel hint can be provided with the input cursor. This indicates that a degree of parallelism up to 4 should be enabled.


select *
from table(churnGroupEval(
  cursor(select * /*+ parallel(t,4) */ from CHURN_TRAIN t),
  cursor(select 1 as "ore.connect",' myC5.0model2' as "datastorePrefix" from dual),
  'XML', 'state', 'myC5.0Function'));
Map Reduce

The “group apply” functionality can be thought of in terms of the map-reduce paradigm where the mapper performs the partitioning by outputting the INDEX value as key and the data.frame as value. Then, each reducer receives the rows associated with one key. In our example above, INDEX was the column state and so each reducer would receive rows associated with a single state.

Memory and performance considerations

While the data is partitioned by the INDEX column, it is still possible that a given partition is quite large, such that either the partition of data will not fit in the R engine memory or the user-defined embedded R function will not be able to execute to completion. The usual remedial measures can be taken regarding setting memory limits – as noted in Part 2.

If the partitions are not balanced, you would have to configure the system’s memory for the largest partition. This will also have implications for performance, obviously, since smaller partitions of data will likely complete faster than larger ones.

The blog post Managing Memory Limits and Configuring Exadata for Embedded R Execution discusses how to instrument your code to understand the memory usage of your R function. This is done in the context of ore.indexApply (to be discussed later in this blog series), but the approach is analogous for “group apply.”

Friday Jan 03, 2014

Invoking R scripts via Oracle Database: Theme and Variation, Part 2

In part 1 of Invoking R scripts via Oracle Database: Theme and Variation, we introduced features of Oracle R Enterprise embedded R execution, focusing on the functions ore.doEval and rqEval. In this blog post, we’ll cover the next in our theme and variation series involving ore.tableApply and rqTableEval.

The variation on embedded R execution for ore.tableApply involves passing an ore.frame to the function such that the first parameter of your embedded R function receives a data.frame. The rqTableEval function in SQL allows users to specify a data cursor to be delivered to your embedded R function as a data.frame.

Let’s look at a few examples.


R API

In the following example, we’re using ore.tableApply to build a Naïve Bayes model on the iris data set. Naïve Bayes is found in the e1071 package, which must be installed on both the client and database server machine R engines.

library(e1071)
mod <- ore.tableApply(
ore.push(iris),
function(dat) {
library(e1071)
dat$Species <- as.factor(dat$Species)
naiveBayes(Species ~ ., dat)
})
class(mod)
mod

A few points to highlight:
• To use the CRAN package e1071 on the client, we first load the library.
• The iris data set is pushed to the database to create an ore.frame as the first argument to ore.tableApply. This would normally refer to an ore.frame that refers to a table that exists in Oracle Database. If not obvious, note that we could have previously assigned dat <- ore.push(iris) and passed dat as the argument as well.
• The embedded R function is supplied as the second argument to ore.tableApply as a function object. Recall from Part 1 that we could have alternatively assigned this function to a variable and passed the variable as an argument, or stored the function in the R script repository and passed the argument FUN.NAME with the assigned function name.
• The user-defined embedded R function takes dat as its first argument which will contain a data.frame derived from the ore.frame supplied.
• The model itself is returned from the function.
• The result of the ore.tableApply execution will be an ore.object.

SQL API

We can invoke the function through the SQL API by storing the function in the R script repository. Recall that the call to sys.rqScriptCreate must be wrapped in a BEGIN-END PL/SQL block.


begin
sys.rqScriptCreate('myNaiveBayesModel',
'function(dat) {
library(e1071)
dat$Species <- as.factor(dat$Species)
naiveBayes(Species ~ ., dat)
}');
end;
/

Invoking the function myNaiveBayesModel occurs in a SQL SELECT statement as shown below. The first argument to rqTableEval specifies a cursor that retrieves the IRIS table. Note that the IRIS table could have been created earlier using ore.create(iris,"IRIS"). The second argument, NULL, indicates that no arguments are supplied to the function.

The function returns an R object of type naiveBayes, but as a serialized object that is chunked into a table. This likely is not useful to most users.


select *
from table(rqTableEval(cursor(select * from IRIS), NULL, NULL, 'myNaiveBayesModel'));

If we want to keep the model in a more usable form, we can store it in an ORE datastore in Oracle Database. For this, we require a change to the user-defined R function and the SQL invocation.


begin
sys.rqScriptCreate('myNaiveBayesModel',
'function(dat) {
library(e1071)
dat$Species <- as.factor(dat$Species)
mod <- naiveBayes(Species ~ ., dat)
ore.save(mod, name="myNaiveBayesDatastore")
TRUE

}');
end;
/
select *
from table(rqTableEval(cursor(select * from IRIS), cursor(select 1 as "ore.connect" from dual), 'XML', 'myNaiveBayesModel'));

Highlighted in red, we’ve stored the model in the datastore named ‘myNaiveBayesDatastore’. We’ve also returned TRUE to have a simple value that can show up as the result of the function execution. In the SQL query, we changed the third parameter to ‘XML’ to return an XML string containing “TRUE”. The name of the datastore could be passed as an argument as follows:


begin
sys.rqScriptCreate('myNaiveBayesModel',
'function(dat, datastoreName) {
library(e1071)
dat$Species <- as.factor(dat$Species)
mod <- naiveBayes(Species ~ ., dat)
ore.save(mod, name=datastoreName)
TRUE
}');
end;
/
select *
from table(rqTableEval(
cursor(select * from IRIS),
cursor(select 'myNaiveBayesDatastore' "datastoreName", 1 as "ore.connect" from dual),
'XML',
'myNaiveBayesModel'));

Memory considerations with ore.tableApply and rqTableEval

The input data provided as the first argument to a user-defined R function invoked using ore.tableApply or rqTableEval is physically being moved from Oracle Database to the database server R engine. It’s important to realize that R’s memory limitations still apply. If your database server machine has 32 GB RAM and your data table is 64 GB, ORE will not be able to load the data into the R function’s dat argument.
You may see errors like:


Error : vector memory exhausted (limit reached)

or

ORA-28579: network error during callback from external procedure agent

See the blog post on Managing Memory Limits and Configuring Exadata for Embedded R Execution where we discuss setting memory limits for the database server R engine. This can be necessary to load reasonably sized data tables.

Parallelism

As with ore.doEval / rqEval, user-defined R functions invoked using ore.tableApply / rqTableEval are not executed in parallel, i.e., a single R engine is used to execute the user-defined R function.

Invoking certain ORE advanced analytics functions

In the current ORE release, some advanced analytics functions, like ore.lm or ore.glm, which use the embedded R execution framework, cannot be used within other embedded R calls such as ore.doEval / rqEval and ore.tableApply / rqTableEval.

You can expect to see an error like the following:


ORA-28580: recursive external procedures are not supported

In the next post in this series, I’ll discuss ore.groupApply and the corresponding definitions required for SQL execution, since there is no rqGroupApply function. I’ll also cover the relationship of various “group apply” constructs to map-reduce paradigm.

ORAAH - Enabling high performance R workloads on Hadoop


One of the features of Oracle R Advanced Analytics for Hadoop (ORAAH) is enabling Hadoop jobs written in the R language. R is a popular open-source language and environment for statistical computing and graphics. ORAAH enables R programmers to leverage a Hadoop cluster operating on data resident in HDFS files.

In this blog post, we examine the performance characteristics of ORAAH with an example and explain what makes ORAAH the fastest alternative available to run Hadoop-R jobs. We also compare the results with another popular Hadoop interface for R, rmr.

Credit to Vlad Sharanhovich and Anand Srinivasan for providing the content for this blog post.

In probability theory and statistics, covariance is a measure of how much two variables change together.

Variables that tend to show similar behavior exhibit positive covariance. Alternatively if the greater values of one variable correspond with the smaller values of another then the covariance between the variables is negative. We use covariance computation as the running example below that you can use to reproduce the results detailed here.

The tests were performed on a 6-node cluster running ORAAH version 2.3.1.


Cluster configuration:


  • 6 node cluster

  • BDA v2.3.1 (based on CDH 4.4)

  • 4 tasktrackers

  • CPU: Intel ® Xeon® CPU X5675  @ 3.07GHz

  • RAM: 47GB


ORAAH comes with a convenience function for data generation. We use this function to generate a 100 GB HDFS input file with numeric values for 200 variables (columns), as follows:


# Generate 100GB input dataset (using pre-release 2.4.0)
inputCsv <- orch.datagen(1e+11, numeric.col.count=200, parts=100)
# 15 mins, 53 sec
# 200 mappers

ORAAH supports 2 types of HDFS input: delimited text files and a binary RDATA representation (R's own binary representation). In many cases, RDATA representation provides much better I/O throughput compared to delimited text files.


# Converting into ORAAH native format
inputRdata <- hdfs.toRData(inputCsv, out.name="100G_200n_rd")
# 4 mins, 54 sec
# 400 mappers

Next, we write the mapper and reducer code for parallel/distributed covariance computation. Below, the mapper function accepts a data.frame representation of the input data and generates structured output with 3 components: a matrix, a vector of column sums, and input row count. The single reducer function merges the structured output generated from the mappers to produce the final covariance matrix.

The mapper and reducer functions are supplied as input to ORAAH's hadoop.run() function. This function additionally takes an HDFS file as input in the dataargument. The structured output from the mapper and reducer is defined in mapred.config data structure. Optionally, the Hadoop job can be given a name (in this case "cov") for traceability.

ORCH_cov
<- function(x) {
hadoop.run(
data = x,
mapper = function(k, v) {
m <- as.matrix(v)
cs <- colSums(m)
nr <- nrow(m)
mtm <- t(m) %*% m
l <- list(mat=mtm, colsum=cs, nrow=nr)
orch.keyval(NULL, orch.pack(l)),
reducer = function(k, v) {
mapres <- orch.unpack(v$val, as.list=T)
xy <- Reduce("+", lapply(mapres,function(x) x$mat))
csf <- Reduce("+", lapply(mapres,function(x) x$colsum))
nrf <- Reduce("+", lapply(mapres,function(x) x$nrow))
sts <- csf %*% t(csf)
m1 <- xy / (nrf -1)
m2 <- sts / (nrf * (nrf-1))
m3 <- 2 * sts / (nrf * (nrf-1))
covmat <- m1 + m2 - m3
orch.keyval(NULL, orch.pack(covmat))
},
config = new("mapred.config",
map.output = data.frame(key="none",
val="c"),
reduce.output = data.frame(key="none", val="c")),
job.name = "cov")
}

The execution times are shown below:


# Using text input
ORCH_cov(inputCsv)
# 7 mins, 19 sec
# 400 mappers / 45 reducers



# Using binary input
ORCH_cov(inputRdata)
# 4 mins, 18 sec
# 400 mappers / 45 reducers

A few points to observe:
1. Before an HDFS file can be used with ORAAH's hadoop.run() function, its metadata must be known. ORAAH automatically determines the data types of the columns in the HDFS file by sampling rows. The metadata is created during the hdfs.attach() call. This metadata enables ORAAH to generate highly optimized scan routines to read rows from the file.
2. ORAAH implements caching of input and output structures from mappers and reducers thereby lifting the burden on dealing with large data volumes from the R engine.
3. ORAAH's orch.pack() and orch.unpack() functions enable transfer of structured constructs between mappers and reducers, which further improves I/O throughput by eliminating the need to scan/parse string inputs.
4. ORAAH leverages R's own RDATA representation as the binary representation. The key to better execution performance of R jobs on Hadoop is managing I/O throughput and carefully bypassing R's inherent limitations with parsing strings.

We contrast ORAAH's performance by comparing it with an open source package called rmr (https://github.com/RevolutionAnalytics/RHadoop/wiki/rmr) repeating the covariance calculation on the same cluster and input data set. For this experiment, we used rmr version 2.3.0.

The covariance computation is written in rmr as shown below.


RMR2_cov <- function(x, input.format) {
mapreduce(
x,
input.format = input.format,
map = function(k, v) {
m <- as.matrix(v)
cs <- colSums(m)
nr <- nrow(m)
mtm <- t(m) %*% m
l <- list(mat=mtm, colsum=cs, nrow=nr)
keyval(1, paste0(deparse(l),
collapse="\n"))
},
reduce = function(k, v) {
mapres <- lapply(v, function(x)
eval(parse(text=x)))
xy <- Reduce("+", lapply(mapres, function(x) x$mat))
csf <- Reduce("+", lapply(mapres, function(x) x$colsum))
nrf <- Reduce("+", lapply(mapres, function(x) x$nrow))
sts <- csf %*% t(csf)
m1 <- xy / (nrf -1)
m2 <- sts / (nrf * (nrf-1))
m3 <- 2 * sts / (nrf * (nrf-1))
covmat <- m1 + m2 - m3
keyval(1, paste0(deparse(covmat),
collapse="\n"))
}
)
}

Notice that the code is somewhat similar to the version used with ORAAH with 3 key differences:
   i) Mapper output is serialized as a string
 ii) Reducer, thus, is forced to parse input strings
iii) Reducer output is once again a string, requiring client to reconstruct the covariance matrix

Not only is passing strings limiting from the R programmer's perspective, but also has a negative effect on I/O throughput. rmr also supports a proprietary binary representation of delimited text data.

Below we repeat the tests with both delimite and binary representations. Not only is the conversion to the binary representation more expensive, the resulting I/O throughput is not substantially better.


# Convert to RMR2 native format
inputRMR <- mapreduce(inputCsv, 
  input.format = make.input.format("csv",sep=","),
  map = function(k, v) keyval(NULL, v)
)
# 20 mins, 17 sec
# 400 mappers
#
# Using text input
RMR2_cov(inputCsv, make.input.format("csv",sep=","))
# 32 mins, 14 sec
# 400 mappers / 45 reducers
#
# Using binary input
RMR2_cov(inputRMR, "native")
# 17 mins, 18 sec
# 400 mappers / 45 reducers

To summarize ORAAH is 4x faster than rmr out of the box for a simple covariance calculation.





























 Text Input
 Binary Input
 Text to Binary Conversion
 ORAAH  7 min, 19 sec
 4 min, 18 sec
 4 min 54 sec
 rmr  32 min, 14 sec
 17 min, 18 sec
 20 min 17 sec

 4.4x faster
 4x faster
 4.14x faster

Thursday Jan 02, 2014

Invoking R scripts via Oracle Database: Theme and Variation


Oracle R Enterprise provides several ways for you to invoke R scripts through Oracle Database. From the same R script you can get structured data, an XML representation of R objects and images, and even PNG images via a BLOB column in a database table. This series of blog posts will take you through the various ways you can interact with R scripts and Oracle Database. In this first post, we explore the benefits of embedded R execution and usage of the ore.doEval and rqEval functions. Subsequent posts will detail the use of the other embedded R execution functions, and in the case of data- and task-parallel capabilities, how these relate to the MapReduce paradigm.

Introduction


In Oracle R Enterprise, we use the phrase “embedded R execution” to characterize the storing of R scripts in Oracle Database – using the ORE R script repository – and invoking that script in one or more database-side R engines.

This is a powerful capability for several reasons:


  • enable data- and task-parallel execution of user-defined R functions that correspond to special cases of Hadoop Map-Reduce jobs

  • leverage a more powerful database server machine for R engine execution – both RAM and CPU

  • transfer data between Oracle Database and R engine much faster than to a separate client R engine

  • invoke user-defined R functions from SQL and retrieve results in various forms depending on application requirements: tables, XML, PNG BLOB

  • leverage open source CRAN packages at the database server machine

  • schedule R scripts for automatic execution via SQL with Oracle Database DBMS_SCHEDULER PL/SQL package


Users can interactively develop R scripts using their favorite R IDE, and then deploy the script as an R function to the database where it can be invoked either from R or SQL. Embedded R execution facilitates application use of R scripts with better performance and throughput than using a client-side R engine. Executing R scripts from SQL enables integration of R script results with OBIEE, Oracle BI Publisher, and other SQL-enabled tools for structured data, R objects, and images.

Table 1 provides a summary of the embedded R execution functions and R script repository functions available. The function f refers to the user-defined R code, or script, that is provided as either an R function object or a named R function in the database R script repository. To create functions in the R script repository, ORE has functions as described in Table 1.












































R API SQL API Description

ore.doEval

rqEval

Executes f with no automatic transfer of data.

ore.tableApply

rqTableEval

Executes f passing all rows of provided input ore.frame as the first parameter of f. First parameter provided as a data.frame.

ore.groupApply

“rqGroupEval”
(must
be explicitly defined as function by user)

Executes f by partitioning data according to an “index” column’s values. Each data partition provided as a data.frame in the first parameter of f. Supports parallel execution of each f invocation in a pool of database server-side R engines.

ore.rowApply

rqRowEval

Executes f passing a specified number of rows (a “chunk”) of the provided input ore.frame. Each chunk provided as a data.frame in the first parameter of f. Supports parallel execution of each f invocation in a pool of database server-side R engines.

ore.indexApply

N/A

Executes f with no automatic transfer of data, but provides the index of the invocation, 1 through n, where n is the number of functions to invoke. Supports parallel execution of each f invocation in a pool of database server-side R engines.

ore.scriptCreate

sys.rqScriptCreate

Load the provided R function into the R script repository with the provided name.

ore.scriptDrop

sys.rqScriptDrop

Remove the named R function from the R script repository.

Table 1: Embedded R Execution and R Script Repository function summary


Using ore.doEval and rqEval


The first of the embedded R functions we cover are also the simplest. The R function ore.doEval and the SQL function rqEval do not automatically receive any data from the database. They simply execute the function f provided. Any needed data is either generated within f or explicitly retrieved from a data source such as Oracle Database, other databases, or flat files.

R API


In the R interface, users can specify an R function as an argument to ore.doEval, or use the name of that function as stored in the R script repository. For example, the function RandomRedDots returns a data.frame with two columns and plots 100 random normal values. To invoke the function through the database server requires minimal specification with ore.doEval.

RandomRedDots <- function(divisor=100){
id<- 1:10
plot(1:100, rnorm(100), pch = 21, bg = "red", cex = 2 )
data.frame(id=id, val=id / divisor)
}
ore.doEval(RandomRedDots)


Here is the result, where the image is displayed at the client, but generated by the database server R engine that executed the function f.


We can provide arguments to f as well. To override the divisor argument, provide it as an argument to ore.doEval. Note
that any number of parameters, including more complex R objects such as models can be passed as arguments this way.

ore.doEval(RandomRedDots, divisor=50)


Behind the scenes: when passing the function itself (as above), ORE implicitly stores the function in the R script repository before executing it. When finished executing, the function is dropped from the repository. If we want to store this function explicitly in the repository, we can use ore.scriptCreate:

ore.scriptCreate("myRandomRedDots",
RandomRedDots)


Now, the function can be invoked by name:

ore.doEval(FUN.NAME="myRandomRedDots",divisor=50)


The return value of f is a data.frame, however, if we capture the result in a variable, we’ll notice two things: the class of the return value is ore.object and the image does not display.

res <- ore.doEval(FUN.NAME="myRandomRedDots",
divisor=50)
class(res)



To get back the data.frame, we must invoke ore.pull to pull the result to the client R engine.

res.local <- ore.pull(res)
class(res.local)


If we wanted to return an ore.frame instead of an ore.object, specify the argument FUN.VALUE that describes the structure of the result.

res.of <- ore.doEval(FUN.NAME="myRandomRedDots", divisor=50,
FUN.VALUE= data.frame(id=1, val=1))
class(res.of)


SQL API


Now, we’ll look at executing the same R function f using the SQL interface of embedded R execution, while pointing out a few significant differences in the API. The first difference is that the R functions are defined as strings, not R objects. This should be no surprise since we’ll be using a SQL interface like SQL Developer or SQL*Plus. Also, the R function string cannot be passed directly in the rqEval function, but must first be stored in the R script repository. The call to sys.rqScriptCreate must be wrapped in a BEGIN-END PL/SQL block.

begin
sys.rqScriptCreate('myRandomRedDots2',
'function(divisor=100,numDots=100){
id <- 1:10
plot( 1:numDots, rnorm(numDots), pch = 21, bg = "red", cex = 2 )
data.frame(id = id, val = id / divisor)
}');
end;
/


Invoking the function myRandomRedDots2 occurs in a SQL SELECT statement as shown below. The first NULL argument to rqEval indicates that no arguments are supplied to the function myRandomRedDots2. In the SQL API, we can ask for the data.frame returned by f to appear as a SQL table. For this, the second parameter can take a SQL string that describes the column names and data types that correspond to the returned data.frame. You can provide a prototype row using the dual dummy table, however, the select statement can be based on an existing table or view as well.

select *
from table(rqEval(NULL, 'select 1 id, 1 val from dual', 'myRandomRedDots2'));


To pass parameters in SQL, we can replace the first NULL argument with a cursor that specifies a single row of scalar values. Multiple arguments can be specified as shown below. Note that argument names are case sensitive, so it is best to include column names in double quotes. Note also that the first argument is a cursor whereas the second parameter is a string. The former provides data values, whereas the latter is parsed to determine the structure of the result.

select *
from table(rqEval(cursor(select 50 "divisor", 500 "numDots" from dual),
'select 1 id, 1 val from dual',
'myRandomRedDots2'));


When specifying a table structure for the result as above, any image data is discarded. To get back both structured data and images, we replace the second argument with ‘XML’. This instructs the database to generate an XML string, first with any structured or semi-structured R objects, followed by the image or images generated by the R function f. Images are returned as a base 64 encoding of the PNG representation.

select *
from table(rqEval(cursor(select 50 "divisor", 500 "numDots" from dual),
'XML',
'myRandomRedDots2'));


Advanced features


To establish a connection to Oracle Database within the R function f, a special argument ore.connect can be set to TRUE. This uses the credentials of the user who invoked the embedded R function ore.doEval or rqEval to establish a connection and also automatically load the ORE package. This capability can be useful to explicitly use the ORE Transparency Layer or to save and load objects with ORE R object datastores.

RandomRedDots <- function(divisor=100, datatstore.name="myDatastore"){
id <- 1:10
plot(1:100, rnorm(100), pch = 21, bg = "red", cex = 2 )
ore.load(datastore.name) # contains numeric variable myVar
data.frame(id=id, val=id / divisor, num=myVar)
}


ore.doEval(RandomRedDots, datastore.name="datastore_1", ore.connect=TRUE)


Notice the additions in red. We pass the name of a datastore to load. That datastore is expected to contain a variable myVar. Arguments prefixed with ‘ore.’ are control arguments and are not passed to f. Other control arguments include: ore.drop which if set to TRUE converts a one-column input data.frame to a vector, ore.graphics which if set to TRUE starts a graphical driver to look for images being returned from f, ore.png.* which provides additional parameters for the PNG graphics device. The ore.png.* control arguments include (replace * with): width, height, units, pointsize, bg, res, type, etc.

In the next post, we will explore ore.tableApply and rqTableEval.

Wednesday Dec 04, 2013

Using DCLI to install Oracle R Distribution and Oracle R Enterprise

Oracle R Enterprise is commonly used to apply parallel resources to R computations in Oracle's Exadata Database Machine, To take advantage of Exadata's massively parallel grid infrastructure, Oracle R Distribution and the Oracle R Enterprise server components must be installed on each node. We've now streamlined the installation of Oracle R on Exadata, allowing users to get up and running quickly.

This is where Exadata's distributed command line interface utility (DCLI) comes in handy - it can be used to control multiple nodes with a single command. In Exadata environments, it's common to use DCLI to manage or monitor multiple nodes simultaneously to eliminate having to log in to each node individually.  In this post, we will use DCLI to install Oracle R Distribution and Oracle R Enterprise server components on Exadata compute nodes.

DCLI comes with a help flag that indicates the various options and commands.  We will use some of these commands in the following steps. The flags for non-Exadata RAC systems may differ slightly, so these instructions may require slight modifications for non-Exadata RAC environments. Refer to My Oracle Support for assistance with DCLI options.

$ dcli -h

Distributed Shell for Oracle Storage

This script executes commands on multiple cells in parallel threads.
The cells are referenced by their domain name or ip address.
Local files can be copied to cells and executed on cells.
This tool does not support interactive sessions with host applications.
Use of this tool assumes ssh is running on local host and cells.
The -k option should be used initially to perform key exchange with
cells.  User may be prompted to acknowledge cell authenticity, and
may be prompted for the remote user password.  This -k step is serialized
to prevent overlayed prompts.  After -k option is used once, then
subsequent commands to the same cells do not require -k and will not require
passwords for that user from the host.
Command output (stdout and stderr) is collected and displayed after the
copy and command execution has finished on all cells.
Options allow this command output to be abbreviated.

Return values:
 0 -- file or command was copied and executed successfully on all cells
 1 -- one or more cells could not be reached or remote execution
 returned non-zero status.
 2 -- An error prevented any command execution

Examples:
 dcli -g mycells -k
 dcli -c stsd2s2,stsd2s3 vmstat
 dcli -g mycells cellcli -e alter iormplan active
 dcli -g mycells -x reConfig.scl

usage: dcli [options] [command]

options:
 --version           show program's version number and exit
 -c CELLS            comma-separated list of cells
 -d DESTFILE         destination directory or file
 -f FILE             file to be copied
 -g GROUPFILE        file containing list of cells
 -h, --help          show help message and exit
 -k                  push ssh key to cell's authorized_keys file
 -l USERID           user to login as on remote cells (default: celladmin)
 -n                  abbreviate non-error output
 -r REGEXP           abbreviate output lines matching a regular expression
 -s SSHOPTIONS       string of options passed through to ssh
 --scp=SCPOPTIONS    string of options passed through to scp if different
 from sshoptions
 --serial            serialize execution over the cells
 -t                  list target cells
 --unkey             drop keys from target cells' authorized_keys file
 -v                  print extra messages to stdout
 --vmstat=VMSTATOPS  vmstat command options
 -x EXECFILE         file to be copied and executed

I. Install Oracle R Distribution across Exadata compute nodes

Oracle R Distribution is distributed as a set of RPMs. Root or sudo access is required only to install the RPMs, which is typical of RPM-based installs.  Root access is not necessary for running the R software - no R process will ever run as root.  The Oracle R Enterprise server installation steps are executed by user oracle, or any database user that meets the requirements listed in the Oracle R Enterprise Installation and Administration Guide. We strongly recommend reviewing the prerequisites and installation steps in the documentation prior to beginning the installation.

Step 1: SSH Trust and User Equivalence

The first task is to establish trust between your hosts. In other words, configure the Exadata environment to enable automatic authentication as DCLI executes remote commands.

 a. Generate an SSH public-private key for the root user, on any compute node, as
root.

$ ssh-keygen -N '' -f ~/.ssh/id_dsa -t dsa

This places the generated public and private key files in the .ssh sub-directory of the root user's home directory.

 b. Using your text editor, create a file that contains the host or node names of all the compute nodes in the rack separated by newlines. For example, the nodes file for a 2-node cluster may contain following entries:

$ cat nodes
exadb01
exadb02

c. Run the DCLI command with the -k option to push the SSH public key to each compute node's ssh authorized key file to establish SSH Trust. You will be prompted to enter the password for each compute node, but this will be the only time. With the -k option, each compute node is contacted sequentially rather than in parallel to give you a chance to enter the password for each node.

$ dcli -t -g nodes -l root -k -s "\-o StrictHostkeyChecking=no"

Once the DCLI command completes, you have established SSH Trust and User Equivalence. Subsequent DCLI commands issued will be executed without being prompted for passwords.

Step 2: Log in as root to any compute node in the Exadata and download the file ord-linux-x86_64-version.tar.gz, where version is the R version you want to install. For example, the file name for R-2.15.3 is ord-linux-x86_64-2.15.3.tar.gz.

Step 3: Create a directory and replicate the file ord-linux-x86_64-2.15.3.tar.gz in this directory across all nodes. Here, we create a directory named ORD in /home/oracle and replicate ord-linux-x86_64-2.15.3.tar.gz in /home/oracle/ORD.

$ dcli -t -g nodes -l root mkdir -p /home/oracle/ORD
$ dcli -t -g nodes -l root -f ord-linux-x86_64-2.15.3.tar.gz -d /home/oracle/ORD/ord-linux-x86_64-2.15.3.tar.gz

Step 4: Uncompress and untar ord-linux-x86_64-2.15.3.tar.gz to get the dependent RPMs across all nodes.

$ dcli -t -g nodes -l root tar xvfz /home/oracle/ORD/ord-linux-x86_64-2.15.3.tar.gz -C /home/oracle/ORD
$ ls /home/oracle/ORD/ord-linux-x86_64-2.15.3

NOTE: You can also download these RPMs from
           http://public-yum.oracle.com/

At the time of this blog post, several of these dependencies required by R's development RPMs will cause conflicts during standard Exadata upgrades. To avoid this, remove gcc-gfortran, mesa-libGl-devel, libpng-devel, and R-devel-<version>.el5.x86_64.rpm from the list. For Oracle R Distribution 2.15.3, the RPM is R-devel-2.15.3-1.el5.x86_64.rpm.

Step 5: To install these new RPMs and update existing RPMs across nodes, issue the following RPM command.

$ dcli -t -g nodes -l root rpm -i --force /home/oracle/ORD/ord-linux-x86_64-2.15.3/*.rpm

The --force flag is used here to avoid errors regarding circular dependencies. 

Step 6: Verify R installations on each node by first returning the location where R is installed and then starting R.

$ dcli -g nodes -l oracle R RHOME
$ dcli -g nodes -l oracle R --vanilla

Oracle R Distribution installation on Exadata commands summary:

ssh-keygen -N " -f~/.ssh/id_dsa -t dsa

vi nodes # enter node names

dcli -t -g nodes -l root -k -s "\-o StrictHostkeyChecking=no" 

dcli -t -g nodes -l root mkdir -p /home/oracle/ORD
dcli -t -g nodes -l root -f ord-linux-x86_64-2.15.3.tar.gz -d /home/oracle/ORD/ord-linux-x86_64-2.15.3.tar.gz 
dcli -t -g nodes -l root tar xvfz /home/oracle/ORD/ord-linux-x86_64-2.15.3.tar.gz -C /home/oracle/ORD

dcli -t -g nodes -l root rpm -i --force /home/oracle/ORD/ord-linux-x86_64-2.15.3/*.rpm

dcli -g nodes -l root R RHOME
dcli -g nodes -l root R --vanilla


II. Install Oracle R Enterprise Server across Exadata compute nodes

Before installing Oracle R Enterprise Server, ensure that environment variables are set on each node as shown in Table 4-1 in the Oracle R Enterprise Installation and Administration Guide. 

Step 1: Download Oracle R Enterprise server and supporting installers.

On the Oracle R Enterprise page, download Oracle R Enterprise Server and Oracle R Enterprise client supporting packages for Linux. The following files are downloaded for Oracle R Enterprise 1.3.1:

ore-server-linux-x86-64-1.3.1.zip
ore-supporting-linux-x86-64-1.3.1.zip

Step 2: Copy the Oracle R Enterprise server and supporting packages installers across nodes.

$ dcli -g nodes -l oracle mkdir -p /home/oracle/ORE
$ dcli -g nodes -l oracle -f ore-server-linux-x86-64-1.3.1.zip -d /home/oracle/ORE/ore-server-linux-x86-64-1.3.1.zip
$ dcli -g nodes -l oracle -f ore-supporting-linux-x86-64-1.3.1.zip -d /home/oracle/ORE/ore-supporting-linux-x86-64-1.3.1.zip

Step 3: Unzip Oracle R Enterprise server and supporting packages installers.

$ dcli -t -g nodes -l oracle unzip /home/oracle/ORE/ore-server-linux-x86-64-1.3.1.zip -d /home/oracle/ORE
$ dcli -t -g nodes -l oracle unzip /home/oracle/ORE/ore-supporting-linux-x86-64-1.3.1.zip -d /home/oracle/ORE

Step 4: Install Oracle R Enterprise server components.

$ dcli -t -g nodes -l oracle /home/oracle/ORE/server/./install.sh

Step 5: Create an Oracle R Enterprise user.

$ dcli -t -g nodes -l oracle /home/oracle/ORE/server/./demo_user.sh

Step 6: Apply grants to the Oracle R Enterprise user by executing the demo_user.sh script.  The default user is rquser. As this is a shared database, the grants need only be applied to a single node.

$ cd /home/oracle/ORE 
$ sqlplus / as sysdba

SQL> grant RQADMIN to rquser;
SQL> grant CREATE TABLE to rquser;
SQL> grant CREATE SESSION to rquser;
SQL> grant CREATE VIEW to rquser;
SQL> grant CREATE PROCEDURE to rquser;
SQL> grant CREATE MINING MODEL to rquser;

Step 7: Install Oracle R Enterprise client supporting packages.

$ dcli -t -g nodes -l oracle R CMD INSTALL /home/oracle/ORE/supporting/DBI_0.2-5_R_x86_64-unknown-linux-gnu.tar.gz
$ dcli -t -g nodes -l oracle R CMD INSTALL /home/oracle/ORE/supporting/ROracle_1.1-9_R_x86_64-unknown-linux-gnu.tar.gz
$ dcli -t -g nodes -l oracle R CMD INSTALL /home/oracle/ORE/supporting/png_0.1-4_R_x86_64-unknown-linux-gnu.tar.gz

Step 8: Verify Oracle R Enterprise loads.

$ dcli -t -g nodes -l oracle ORE -e "library(ORE)"

 Additional steps for validating the Oracle R Enterprise installation are in sections 6.3 and 6.4 of the Oracle R Installation and Administration Guide.

Oracle R Enterprise installation on Exadata commands summary:


dcli -g nodes -l oracle mkdir -p /home/oracle/ORE
dcli -g nodes -l oracle -f ore-server-linux-x86-64-1.3.1.zip -d /home/oracle/ORE/ore-server-linux-x86-64-1.3.1.zip
dcli -g nodes -l oracle -f ore-supporting-linux-x86-64-1.3.1.zip -d /home/oracle/ORE/ore-supporting-linux-x86-64-1.3.1.zip

dcli -t -g nodes -l oracle unzip /home/oracle/ORE/ore-server-linux-x86-64-1.3.1.zip -d /home/oracle/ORE
dcli -t -g nodes -l oracle unzip /home/oracle/ORE/ore-supporting-linux-x86-64-1.3.1.zip -d /home/oracle/ORE

dcli -t -g nodes -l oracle /home/oracle/ORE/server/./install.sh
dcli -t -g nodes -l oracle /home/oracle/ORE/server/./demo_user.sh

cd /home/oracle/ORE 
sqlplus / as sysdba

SQL> grant RQADMIN to rquser;
SQL> grant CREATE TABLE to rquser;
SQL> grant CREATE SESSION to rquser;
SQL> grant CREATE VIEW to rquser;
SQL> grant CREATE PROCEDURE to rquser;
SQL> grant CREATE MINING MODEL to rquser;

dcli -t -g nodes -l oracle R CMD INSTALL /home/oracle/ORE/supporting/DBI_0.2-5_R_x86_64-unknown-linux-gnu.tar.gz
dcli -t -g nodes -l oracle R CMD INSTALL /home/oracle/ORE/supporting/ROracle_1.1-9_R_x86_64-unknown-linux-gnu.tar.gz
dcli -t -g nodes -l oracle R CMD INSTALL /home/oracle/ORE/supporting/png_0.1-4_R_x86_64-unknown-linux-gnu.tar.gz

dcli -t -g nodes -l oracle ORE -e "library(ORE)"



Conclusion: DCLI is a powerful utility that provides the ability to install Oracle R Distribution and Oracle R Enterprise on multiple Exadata compute nodes without the effort of repeating commands on each node.


Wednesday Oct 23, 2013

Migrating R Scripts from Development to Production

“How do I move my R scripts stored in one database instance to another? I have my development/test system and want to migrate to production.”

Users of Oracle R Enterprise Embedded R Execution will often store their R scripts in the R Script Repository in Oracle Database, especially when using the ORE SQL API. From previous blog posts, you may recall that Embedded R Execution enables running R scripts managed by Oracle Database using both R and SQL interfaces. In ORE 1.3.1., the SQL API requires scripts to be stored in the database and referenced by name in SQL queries. The SQL API enables seamless integration with database-based applications and ease of production deployment.

Loading R scripts in the repository

Before talking about migration, we’ll first introduce how users store R scripts in Oracle Database. Users can add R scripts to the repository in R using the function ore.scriptCreate, or SQL using the function sys.rqScriptCreate.

For the sample R script

    id <- 1:10
    plot(1:100,rnorm(100),pch=21,bg="red",cex =2)
    data.frame(id=id, val=id / 100)

users wrap this in a function and store it in the R Script Repository with a name. In R, this looks like

ore.scriptCreate("RandomRedDots",
function () {
line-height: 115%; font-family: "Courier New";">     id <- 1:10
    plot(1:100,rnorm(100),pch=21,bg="red",cex =2)
    data.frame(id=id, val=id / 100))
})

In SQL, this looks like

begin
sys.rqScriptCreate('RandomRedDots',

 'function(){
    id <- 1:10
    plot(1:100,rnorm(100),pch=21,bg="red",cex =2)
    data.frame(id=id, val=id / 100)
  }');
end;
/

The R function ore.scriptDrop and SQL function sys.rqScriptDrop can be used to drop these scripts as well. Note that the system will give an error if the script name already exists.

Accessing R scripts once they’ve been loaded

If you’re not using a source code control system, it is possible that your R scripts can be misplaced or files modified, making what is stored in Oracle Database to only or best copy of your R code. If you’ve loaded your R scripts to the database, it is straightforward to access these scripts from the database table SYS.RQ_SCRIPTS. For example,

select * from sys.rq_scripts where name='myScriptName';

From R, scripts in the repository can be loaded into the R client engine using a function similar to the following:

ore.scriptLoad <- function(name) {
query <- paste("select script from sys.rq_scripts where name='",name,"'",sep="")
str.f <- OREbase:::.ore.dbGetQuery(query)
assign(name,eval(parse(text = str.f)),pos=1)
}

ore.scriptLoad("myFunctionName")

This function is also useful if you want to load an existing R script from the repository into another R script in the repository – think modular coding style. Just include this function in the body of the other function and load the named script.

Migrating R scripts from one database instance to another

To move a set of functions from one system to another, the following script loads the functions from one R script repository into the client R engine, then connects to the target database and creates the scripts there with the same names.

scriptNames <- OREbase:::.ore.dbGetQuery("select name from sys.rq_scripts where name not like 'RQG$%' and name not like 'RQ$%'")$NAME

for(s in scriptNames) {
cat(s,"\n")
ore.scriptLoad(s)
}

ore.disconnect()
ore.connect("rquser","orcl","localhost","rquser")

for(s in scriptNames) {
cat(s,"\n")
ore.scriptDrop(s)
ore.scriptCreate(s,get(s))
}

Best Practice

When naming R scripts, keep in mind that the name can be up to 128 characters. As such, consider organizing scripts in a directory structure manner. For example, if an organization has multiple groups or applications sharing the same database and there are multiple components, use “/” to facilitate the function organization:

line-height: 115%;">ore.scriptCreate("/org1/app1/component1/myFuntion1", myFunction1)
ore.scriptCreate("/org1/app1/component1/myFuntion2", myFunction2)
ore.scriptCreate("/org1/app2/component2/myFuntion2", myFunction2)
ore.scriptCreate("/org2/app2/component1/myFuntion3", myFunction3)
ore.scriptCreate("/org3/app2/component1/myFuntion4", myFunction4)

Users can then query for all functions using the path prefix when looking up functions.

Friday Oct 11, 2013

Take Oracle R Enterprise for a Test Drive

If you'd like try Oracle R Enterprise, Oracle Partner Vlamis Software Solutions provides a quick and easy way to get started using a virtual machine (VM) loaded with all the software you require and hosted on Amazon Web Services (AWS).  Follow this link and within a few clicks you'll have a "Remote Desktop" connection to the cloud with sample scripts for you to explore both the R language and Oracle R Enterprise, both from R and SQL.

Oracle R Enterprise, a component of the Oracle Advanced Analytics option, makes the open source R statistical programming language and environment ready for the enterprise and big data. It provides a comprehensive, database-centric environment for end-to-end analytical processes in R, with immediate deployment to production environments. R users can operationalize entire R scripts in production applications, thereby eliminating porting of R code to other languages or reinventing code to integrate R results into existing applications. Oracle R Enterprise allows users to seamlessly leverage Oracle Database as a high performance computing (HPC) environment for R scripts, providing data parallelism and resources management.


Wednesday Oct 02, 2013

Managing Memory Limits and Configuring Exadata for Embedded R Execution

An R engine can consume significant memory resources in the course of running R scripts. R users who work with Oracle R Enterprise Embedded R Execution on sizable data, especially application designers and database administrators (DBAs), have a vested interest in understanding and controlling the memory demands of R script execution to help ensure that sufficient memory resources are available for both their application and Oracle Database. ORE Embedded R Execution enables running R scripts managed by Oracle Database, both through R and SQL APIs. The SQL API enables seamless integration with database-based applications, data-parallel and task-parallel R script execution, and ease of production deployment.

To provide greater control over R memory consumption, Oracle R Enterprise provides a privileged SQL function for configuring a database server with R memory limits. In this blog post, we provide a discussion of R memory usage and garbage collection, and how this SQL function can be used to limit the amount of memory consumed by individual R engines started as part of ORE’s embedded R execution framework. We follow with an example of involving memory limit calculations on Exadata and some recommendations for DBAs to consider when configuring Exadata for embedded R execution. Note that such calculations and configuration settings are applicable to non-Exadata (single instance or custom RAC) environments as well. At the end, there a “tip of the day” for R memory management.

Garbage Collection as a concept

For those familiar with languages like C, memory is explicitly managed by the programmer through invocations of functions to allocate and free memory (malloc, calloc, free). Failing to free memory when finished with it results in “memory leaks” that can cause a process to consume (or exhaust) memory unnecessarily, often resulting in a program or system crash.

To alleviate programmers from this burden, languages like R and Java rely on garbage collection. “Garbage” is memory that is no longer being used, i.e., no longer referenceable, within your program. With garbage collection, programmers avoid dealing with memory management. The underlying system determines what memory is used or available, and frees memory periodically. Garbage collection, however, is not a panacea. Garbage collection can take time to process, e.g., on the order of seconds, which can make response time for certain functions unpredictable – although modern garbage collection mechanisms have largely mitigated this drawback. In addition, when garbage collection occurs is essentially non-deterministic, depending on heuristics set up by the language implementation. This means that memory may be retained longer than necessary.

Memory in R

In R, memory can be characterized along two dimensions: memory allocated for vectors and arrays (referred to as Vcells), and memory allocated for objects such as lists (referred to as “cons” cells or Ncells). When invoking R’s garbage collection function, gc(), you’ll see results like these:

The function gc() returns a matrix with rows Ncells, which corresponds to the cons cells, and Vcells, which corresponds to vector heap memory. The Ncells are 56 bytes/cell (49.2*1024*1024/.920477) on a 64-bit machine, and Vcells are ~8 bytes/cell (22.6*1024*1024/2.956944). The “used” column indicates the number of cells allocated, along with their corresponding megabytes. The column “gc trigger” indicates at what point garbage collection will kick in. The column “max used” indicates the maximum space used since the last call to gc(reset=TRUE) or since R started if gc(reset=TRUE) wasn’t invoked.

As an example of affecting Ncells, consider the following example where we initialize a list as a sequence of 100K numbers. We see that roughly 5.4 MB of RAM were consumed for the 100K cells.

For Vcells, we create a vector of 1M elements. This consumes roughly 3.8 MB of RAM for the 1M cells. R optimizes for integers.

The same test with floats consumes 7.6 MB of RAM for the 1M cells of floats.

How does R’s garbage collector use VSize and NSize?

We’ll discuss VSize, as NSize is analogous. The garbage collector recovers memory that is no longer in use, determining when to perform garbage collection and how much memory to recover. Looking at heap memory for Vcells, as depicted in the figure below, there are a few key points: Min_VSize, VSizeInUse, R VSize, and Max_VSize. The R VSize serves as the gc() trigger. The Min_VSize and Max_VSize are the specified lower and upper memory limits. Min_VSize is the minimal size for the vector heap as well as its initial value. From there, R grows or shrinks the vector heap depending on memory demands. However, it doesn’t exceed the Max_VSize limit nor go below the Min_VSize limit. In the figure, VSizeInUse reflects the memory currently used by R objects. R_VSize is how much memory can be requested without triggering gc(). As you would expect: Min_VSize <= R_VSize <= Max_VSize and VSizeInUse < R_VSize.


Limiting memory on the database server R engine

Oracle R Enterprise provides the SQL function sys.rqconfigset to set memory limits. Use of this function requires the sys privilege and the setting is applied only to embedded R engines. Consider the following examples:

sys.rqconfigset('MIN_VSIZE', '10M') -- min heap 10MB, default 32MB
sys.rqconfigset('MAX_VSIZE', '100M') -- max heap 100MB, default 4GB
sys.rqconfigset('MIN_NSIZE', '500K') -- min number cons cells 500x1024, default 1M
sys.rqconfigset('MAX_NSIZE', '2M') -- max number cons cells 2M, default 20M

Note that either numeric or string values can be provided to sys.rqconfigset. Default constants are defined as follows:


#define RQET_DEF_MINVSZ 33554432 /* RQER DEFault MIN_VSiZe 32Mb */
#define RQET_DEF_MAXVSZ 4294967296 /* RQER DEFault MAX_VSiZe 4Gb */
#define RQET_DEF_MINNSZ 1048576 /* RQER DEFault MIN_NSiZe 1M */
#define RQET_DEF_MAXNSZ 20971520 /* RQER DEFault MAX_NSiZe 20M */

Getting memory settings and usage through an embedded R engine

To obtain the current set of default values, you can invoke the following SQL statement when connected to the database server using the table sys.rq_config.

select name, value from sys.rq_config;

This can be done from within an embedded R function invocation using, for example:

getMemorySettings <- function() {
con <- dbConnect(Extproc())
rs <- dbSendQuery(con, "select name, value from sys.rq_config")
dat <- fetch(rs)
dat
}
ore.doEval(getMemorySettings,ore.connect=TRUE)

To obtain the current memory usage within an individual embedded R engine, instrumenting your embedded R function with gc() and returning the results of gc() will provide this insight:

getMemoryUse <- function() {
gc.dat <- gc()
list(pid=Sys.getpid(), gc.dat=gc.dat)
}
ore.doEval(getMemoryUse)

Note that the result from an embedded R call will also include the memory limits set for the R engines, shown below in column 6 “limit (Mb)” of the result. This occurs whenever memory limits are in place for an R engine. Here, we’ve included getting the process id of the R engine.

An example of computing memory limits

Consider you have an Exadata X2-2 that has 1152 GB RAM (~1.2 TB) and your DBA allocates you a maximum of 60 GB RAM for parallel R engines per Exadata node. If we set the degree of parallelism at 32, to enable 32 R engines to execute concurrently, this allows 1.875 GB RAM / R engine. If we allocate 2/3 of this for Vcells, we would allocate ~1.25 GB for the MAX_VSIZE. The remaining 1/3, or 625 MB, would translate into 11.6M cells for MAX_NSIZE.

60 GB allocated to R engines per Exadata node

DOP=32

60GB / 32 R engines = 1.875 GB / R Engine

~2/3 for Vcells = 1.25 GB; 1.25 GB / 8 Bytes/Cell = 156.25M Cells

~1/3 for Ncells = 625 MB; 625 MB / 54 Bytes/Cell = 11.6M Cells

sys.rqconfigset('MAX_VSIZE', '1250M')
sys.rqconfigset('MAX_NSIZE', '11600K')

While this example focuses on parallel execution, such as for ore.groupApply, ore.rowApply, and ore.indexApply (or SQL rqRowEval and “rqGroupEval”), the same type of analysis applies to non-parallel embedded R functions, like ore.doEval and ore.tableApply (or SQL rqEval and rqTableEval).

Consider an example that builds a randomForest model using the ore.doEval function. We can compute the amount of RAM consumed by the function by invoking gc() at the beginning and end of the function and subtracting the max used “(Mb)” columns as depicted here:

The result is that 1.4 MB were consumed for Ncells and 11.1 MB for Vcells. This can similarly be done for ore.indexApply to see the amount of RAM consumed by each embedded R function execution and to sum up the actual usage for each of the embedded R engines (assuming they run fully concurrently).

Generating such numbers on real data gives users a sense of how much memory embedded R jobs may require.

For DBAs

When configuring a database on Exadata for parallel R engines, consider the following options. In the following scenario, we contrast the scenarios when the execution time of any one given embedded R function is fast, e.g., 10s of seconds, and there are many such executions, versus few parallel R engines where the execution time is long with fewer such executions. Note that these must be considered in context of other Exadata uses:

· Set parallel_degree_policy to MANUAL. This allows ORE to choose when to apply parallelism, as opposed to setting it to AUTO which allows Oracle Database to decide.

· Set parallel_min_servers to the number of parallel slave processes to be started when the database instances start, e.g., 64, which is the number of parallel slave processes per Exadata node. This avoids incurring the time required to start these processes as needed to service R engines, and is particularly important when individual embedded R function execution time is short, e.g., 10s of seconds. If embedded R function execution time is long, the percentage of time for starting up the parallel slave will not dominate the overall execution time.

· Set parallel_max_servers to the maximum number of parallel slave processes that should be allowed per Exadata node, e.g., 128. This ensures that no more than parallel_max_servers will be active at one time, and in turn corresponds to the maximum number of R engines that can be active at one time.

· To avoid overloading the CPUs if the parallel_max_servers limit is reached, set the hidden parameter _parallel_statement_queuing to TRUE. This parameter is turned off by setting parallel_degree_policy to MANUAL. The _parallel_statement_queuing parameter allows for queuing of parallel requests when they exceed the parallel_server_target, which should be set to a value between parallel_min_servers and parallel_max_servers, e.g., 96. Once the parallel_server_target is reached, an embedded R execution will be allowed to execute in parallel using the remaining available parallel servers. If none are available, parallel requests will be queued. This ensures that parallel requests will be run in parallel, as opposed to being forced to serial execution, and be able to take advantage of parallel slaves as they become available. This can dramatically improve overall embedded R execution completion time. Note that parallel_max_servers cannot be changed during database operation, but the parallel_server_target can be to tune Exadata performance. Note that queuing effectively takes no CPU resources.

· To minimize RAC or cluster overhead for fast-executing individual embedded R functions, set parallel_force_local to TRUE to keep all parallel servers allocated and running on the same database server node. With this setting, starting an embedded R execution with DOP 32, all 32 R engines will run on the same Exadata node. A new embedded R execution also with DOP 32 may be started on a different node. If the embedded R functions are long-running, the setup time is propotionately small so spreading the IO over multiple nodes will not adversely impact overall performance. Having parallel slaves span multiple Exadata nodes results in communication / handshaking across nodes, which requires more resources. If the embedded R functions are fast, this overhead can adversely impact overall performance. When all parallel slaves are local, fewer resources are used.

· Where applicable, set application tables and their indexes to DOP 1 to reinforce the ability of ORE to determine when to use parallelism and not be overridden by table or index settings of DEFAULT or a specific degree of parallelism.

R memory management tip

There are many optimizations to make more efficient use of memory. To end this point, here is a tip to reduce memory consumptions significantly and avoid unnecessary replication of data.

If you know the size of your result in advance, pre-allocate the memory required, whether a vector, list, or matrix, as opposed to building up the result incrementally such as using cbind for adding columns to a matrix or data.frame. For example:

num.rows <- 1000
num.cols <- 2000
myFunction <- function(col) {col:(num.rows+col-1)} # produces vector of values
myMatrix <- matrix(NA, num.rows, num.cols) # pre-allocate required memory

for(col in 1:num.cols) {
myMatrix[,col] <- myFunction(col)
}

A note of thanks to Qin Wang and Martin Farber for their input on this blog post.

Monday Aug 12, 2013

Quick! Swap those models – I’ve got a better one

(or, Why in-database analytics enables real-time scoring and can make model deployment easy)

Refreshing predictive models is a standard part of the process when deploying advanced analytics solutions in production environments. In addition, many predictive models need to be used in a real-time setting for scoring customers, whether that is for fraud detection, predicting churn, or recommending next likely product. One of the problems with using vanilla R is that real-time scoring often requires starting an R engine for each score, or enabling some ad hoc mechanism for real-time scoring, which can increase application complexity.

In this blog post, we look at how Oracle R Enterprise enables:

  • Building models in-database on database data from R
  • Renaming in-database models for use by a stored procedure
  • Invoking the stored procedure to make predictions from SQL
  • Building a second model and swapping it with the original
  • Moving a model from development environment to production environment

Building the model in R

So let’s start with building a generalized linear model (GLM) in Oracle Database. For illustration purposes, we’ll use the longley data set from R – a macroeconomic data set that provides a well-known example for a highly collinear regression. In R, type ?longley for the full description of the data set.

Using the following R script, we create the database table LONGLEY_TABLE from the longley data.frame and then build the model using the in-database GLM algorithm. We’re predicting the number of people employed using the remaining variables. Then, we view the model details using summary and the auto-generated fit.name. This fit.name corresponds to the name of the Oracle Data Mining (ODM) model in the database, which is auto-generated. Next, we use the model to predict using the original data, just for a confirmation that the model works as expected.

ore.connect("rquser","my_sid","my_host","rquser_pswd",1521, all=TRUE)

ore.create(longley, table="LONGLEY_TABLE")

mod.glm <- ore.odmGLM(Employed ~ ., data = LONGLEY_TABLE)

summary(mod.glm)

mod.glm$fit.name

predict(fit1, LONGLEY_TABLE)

While a user can refer to the ODM model by its name in fit.name, for example, when working with it in SQL or the Oracle Data Miner GUI, this may not be convenient since it will look something like ORE$23_123. In addition, unless the R object mod.glm is saved in an ORE datastore (an ORE feature corresponding to R’s save and load functions using ore.save and ore.load, but in the database), at the end of the session, this object and corresponding ODM model will be removed.

In addition, we’ll want to have a common name for the model so that we can swap an existing model with a new model and not have the change higher level code. To rename an ODM model, we can use the PL/SQL statement shown here, invoked with R using ore.exec. Of course, this could also be done from any SQL interface, e.g., SQL*Plus, SQL Developer, etc., just supplying the explicit SQL.

ore.exec(paste("BEGIN DBMS_DATA_MINING.RENAME_MODEL(model_name => '", mod.glm$fit.name, "', new_model_name => 'MY_GLM_MODEL'); END;",sep=""))

So now, we have the ODM model named MY_GLM_MODEL. Keep in mind, after the model is renamed, the original model no longer exists and the R object is invalid – at least from the standpoint of being able to use it in functions like summary or predict.

Scoring data from a SQL procedure

As noted above, users can score in batch from R, however, they can also score in batch from SQL. But we’re interested in real-time scoring from the database using the in-database model. This can be done directly in a SQL query but providing the input data in the query itself. This eliminates having to write data to a database table and then doing a lookup to retrieve the data for scoring – making it real-time.

The following SQL does just this. The WITH clause defines the input data, selecting from dual. The SELECT clause uses the model MY_GLM_MODEL to make the prediction using the data defined by data_in.

WITH data_in as (select 2013 "Year",

234.289 "GNP",

235.6 "Unemployed",

107.608 "Population",

159 "Armed.Forces",

83 "GNP.deflator",

60.323 "Employed"

from dual )

SELECT PREDICTION(MY_GLM_MODEL USING *) "PRED"

FROM data_in

While we could invoke the SQL directly, having a stored procedure in the database can give us more flexibility. Here’s the stored procedure version in PL/SQL.

CREATE OR REPLACE

PROCEDURE MY_SCORING_PROC (year_in IN NUMBER,

gnp_in IN BINARY_DOUBLE,

unemployed_in IN BINARY_DOUBLE,

population_in IN BINARY_DOUBLE,

armed_forces_in IN BINARY_DOUBLE,

gnp_deflator_in IN BINARY_DOUBLE,

employed_in IN BINARY_DOUBLE,

pred_out OUT NUMBER) AS

BEGIN

WITH data_in as (select year_in "Year",

gnp_in "GNP",

unemployed_in "Unemployed",

population_in "Population",

armed_forces_in "Armed.Forces",

gnp_deflator_in "GNP.deflator",

employed_in "Employed"

from dual ),

model_score as (SELECT PREDICTION(MY_GLM_MODEL USING *) "PRED"

FROM data_in )

select PRED into pred_out from model_score;

EXCEPTION

WHEN OTHERS THEN

raise_application_error(-20001,

'An error was encountered - '||SQLCODE||' -ERROR- '||SQLERRM);

END;

To invoke the stored procedure, we can do the following:

SET SERVEROUTPUT ON

DECLARE

score NUMBER;

BEGIN

MY_SCORING_PROC(1947, 234.289, 235.6, 107.608, 159, 83, 60.323, score);

DBMS_OUTPUT.PUT_LINE('Score: '|| score);

END;

Refreshing the model from R

Let’s say the model above has been in production for a while, but has become stale – that is, it’s not predicting as well as it used to due to changing patterns in the data. To refresh it, we build a new model. For illustration purposes, we’re going to use the same data (so an identical model will be produced, except for its name).

mod.glm2 <- ore.odmGLM(Employed ~ ., data = LONGLEY_TABLE)

summary(mod.glm2)

mod.glm2$fit.name

To swap the models, we delete the existing model called MY_GLM_MODEL and rename the new model to MY_GLM_MODEL. Again, we can do this from R using PL/SQL and through ore.exec.

ore.exec(paste("BEGIN DBMS_DATA_MINING.DROP_MODEL('MY_GLM_MODEL'); DBMS_DATA_MINING.RENAME_MODEL(model_name => '",mod.glm2$fit.name,"', new_model_name => 'MY_GLM_MODEL'); END;",sep=""))

We can now re-execute the stored procedure and the new model will be used.

SET SERVEROUTPUT ON

DECLARE

score NUMBER;

BEGIN

MY_SCORING_PROC(1947, 234.289, 235.6, 107.608, 159, 83, 60.323, score);

DBMS_OUTPUT.PUT_LINE('Score: '|| score);

END;

You may have noticed that this approach can introduce a brief period where no model is accessible - between the DROP_MODEL and RENAME_MODEL. A better approach involves the use of SYNONYMs. In general, synonyms provide both data independence and location transparency, being an alternative name for a table, view, sequence, procedure, stored function, and other database objects. We can use this in conjunction with our stored procedure above. First, create a synonym for the original scoring procedure.

CREATE or REPLACE SYNONYM MY_SCORING_PROC_SYM for MY_SCORING_PROC;

When invoking the procedure from your application, use the name MY_SCORING_PROC_SYM in place of MY_SCORING_PROC.  Instead of renaming the model, create a second stored procedure, with a different name, e.g., MY_SCORING_PROC_2. The new procedure references the name of the newly build model internally. 

When it is time to swap the models, invoke the following to change the procedures.

 

CREATE or REPLACE SYNONYM MY_SCORING_PROC_SYM for MY_SCORING_PROC_2;

Another benefit of this approach is that replaced models can still be kept should you need to revert to a previous version. 

Moving an in-database model from one machine to another

In a production deployment, there’s often the need to move a model from the development environment to the production environment. For example, the data scientist may have built the model in a development / sandbox environment and now needs to move it to the production machine(s).

In-database models provide functions EXPORT_MODEL and IMPORT_MODEL as part of the DBMS_DATA_MINING SQL package. See the 11g documentation for details. These calls can be invoked from R, but we’ll show this from SQL just to keep the flow easier to see.

From a SQL prompt, e.g., from SQL*Plus, connect to the schema that contains the model. Create a DIRECTORY object where the exported model file will be stored. List the model names available to this schema, which should contain MY_GLM_MODEL. Then, export the model

CONNECT rquser/rquser_psw

CREATE OR REPLACE DIRECTORY rquserdir AS '/home/MY_DIRECTORY';

-- list the models available to rquser

SELECT name FROM dm_user_models;

-- export the model called MY_GLM_MODEL to a dump file in same schema

EXECUTE DBMS_DATA_MINING.EXPORT_MODEL ('MY_GLM_MODEL_out',

'RQUSERDIR',

'name = ''MY_GLM_MODEL''');

At this point, you have the ODM model named MY_GLM_MODEL in the file MY_GLM_MODEL_out01.dmp stored in the file system under /home/MY_DIRECTORY. This file can now be moved to the production environment and the model loaded into the target schema.

Log into the new schema and invoke IMPORT_MODEL.

CONNECT rquser2/rquser2_psw

EXECUTE DBMS_DATA_MINING.IMPORT_MODEL (MY_GLM_MODEL_out01.dmp',

'RQUSERDIR', 'name = ''MY_GLM_MODEL''',

'IMPORT', NULL, 'glm_imp_job', 'rquser:rquser2');

Summary

In this post, we’ve highlighted how to build an in-database model in R and use it for scoring through SQL in a production, re-time settings. In addition, we showed how it is possible to swap, or refresh, models in a way that can leave your application code untouched. Finally, we highlighted database functionality that allows you to move in-database models from one database environment to another.

Users should note that all the functionality shown involving SQL, or being invoked through ore.exec, can be easily wrapped in R functions that could ultimately become part of ORE. If any of our readers are interested in giving this a try, we can post your solution here to share with the R and Oracle community. For the truly adventurous, check out the Oracle Database package DBMS_FILE_TRANSFER to consider wrapping the ability to move model files from R as well.

About

The place for best practices, tips, and tricks for applying Oracle R Enterprise, Oracle R Distribution, ROracle, and Oracle R Advanced Analytics for Hadoop in both traditional and Big Data environments.

Search

Categories
Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today