Monday Aug 04, 2014

Getting started Part 4 - Advanced index customization

This post follows on from Part 1, Part 2 and Part 3 and uses the same example "quickstart" table.

Here we're going to look at some more advanced options for customizing the Oracle Text index.  If you just want basic full-text search on a text column, you may not need to know this at all. However, if you want to maximum use of Oracle Text, it's as well to understand the mechanisms used to customize the index.

So far we've looked at queries run against a default CONTEXT index, which you'll recall we created with the statement:

create index full_name_index on quickstart( full_name ) indextype is ctxsys.context;

CONTEXT indexes have many options available. We can't possibly cover them all here, but instead we'll look at the mechanism for applying customization to an index. This is done by creating a set of preferences, each of which contains its own attributes.  These preferences and attributes are created by calling PL/SQL stored procedures in the package CTX_DDL. It is in order to be able to call these procedures that we granted the CTXAPP role to our user in part one.

First of all, we'll consider defining join characters for our index. We saw earlier that the string "Smith-Smith" was indexed as two separate words, or tokens, in our index.  Indeed, we can check this by peeking at one of the tables which is created automatically when we created the index. The table we need to check is the so-called "$I table" - also known as the token table, which is given the name DR$index_name$I

SQL> select token_text from dr$full_name_index$i;


Breaking up text into words or tokens is handled by a process known as the lexer.  By default, the hyphen character - like all non-alphanumerics charatcers -  is treated by the lexer as a break character. It causes a break between tokens.  We can also change it so that it joins tokens.  It can either become part of the token, so the token is "smith-smith" by defining it as a printjoins character, or it can invisibly join the token as "smithsmith" by defining it as a skipjoins character.  We'll make it a printjoins character.

Looking in the documentation, we find that PRINTJOINS is an attribute of the BASIC_LEXER.  So first we need to create a preference which uses the BASIC_LEXER:

We need to give the preference a name, so we'll call it my_lex_pref:

SQL> execute ctx_ddl.create_preference('my_lex_pref', 'BASIC_LEXER')

Then, for this preference, we need to set the printjoins attribute:

SQL> execute ctx_ddl.set_attribute('my_lex_pref', 'PRINTJOINS', '-')

So now we have a suitable preference, which we can include in the index via an index parameters clause.  First let's drop our existing index:

SQL> drop index full_name_index;

Index dropped.

And create a new one, specifying our newly-created lexer preference in the parameters clause:

SQL> create index full_name_index on quickstart( full_name ) indextype is ctxsys.context
  parameters ('lexer my_lex_pref');

Index created.

Now if we peek into the index table again, we'll see something different:

SQL> select token_text from dr$full_name_index$i;


So we can see that now the "-" character has been included in the indexed token.  We can also test this with a query. By default the "-" character is an operator in the CONTAINS query language (it means the same as MINUS - subtract the score of expression2 from expression1, similar but not the same as NOT), so we will need to escape it to keep it as part of the query:

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'smith\-smith', 99) > 0 order by score(99) desc;

 SCORE(99)         ID FULL_NAME
---------- ---------- ------------------------------
         4          3 Peter Smith-Smith

We've examined one way of customizing the CONTEXT index. There are many other ways that we don't have space to consider here - see the documentation for more details.  Most of these customizations use the same model of creating a preference then setting an attribute for that preference - but there are a couple of exceptions which we'll look at.

Firstly, Oracle Text has the concept of a default stoplist. This contains a list of common "noise" words which are not particularly useful for searching and take up a lot of space in the index.  In English, examples are in, on, and, for, to, that, etc. If we chose, we can use an empty stoplist (predefined as CTXSYS.EMPTY_STOPLIST) or create our own custom stoplist, and add stopwords to it:

SQL> execute ctx_ddl.create_stoplist('my_stoplist')
PL/SQL procedure successfully completed.

SQL> execute ctx_ddl.add_stopword('my_stoplist', 'JOHN')
PL/SQL procedure successfully completed.

SQL> execute ctx_ddl.add_stopword('my_stoplist', 'PETER')
PL/SQL procedure successfully completed.

SQL> drop index full_name_index;
Index dropped.

SQL> create index full_name_index on quickstart( full_name ) indextype is ctxsys.context parameters ('stoplist my_stoplist');
Index created.

Now if we look in our $I table we will see a reduced set of tokens:

SQL> select token_text from dr$full_name_index$i;


So what happens if we use a stopword in a query? The answer is that it matches any word:

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'john smith', 99) > 0 order by score(99) desc;

SCORE(99)         ID FULL_NAME
---------- ---------- ------------------------------
       7            3 Peter Smith-Smith
       4            2 John Smith

The other customization which doesn't use preferences is section groups. These are designed to allow us to search within specific sections of the document such as title, author, etc. The simplest one to use is AUTO_SECTION_GROUP which automatically identifies any section marked up with XML-like tags.  To illustrate this, we'll need to add a new record to our table.

insert into quickstart values (7, 'US', '<first>John</first> <last>Smith</last>');

Then we'll create a new section group using AUTO_SECTION_GROUP, and recreate the index using it:

SQL> execute ctx_ddl.create_section_group('my_sec_grp', 'AUTO_SECTION_GROUP') PL/SQL procedure successfully completed.

SQL> drop index full_name_index;

Index dropped.

SQL> create index full_name_index on quickstart( full_name ) indextype is ctxsys.context parameters ('section group my_sec_grp');

Index created.

And we can then query in the FIRST section using the WITHIN operator:

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'john WITHIN first', 99) > 0 order by score(99) desc;

SCORE(99)          ID FULL_NAME
---------- ---------- ----------------------------------------
         3          7 <first>John</first> <last>Smith</last>

That concludes our brief look at index customization.  There is much more to it - please look through the doc for more ideas.

Getting started Part 3 - Index maintenance

This post follows on from Part 1 and Part 2, and uses the same example "quickstart" table.

One thing that surprised new users is that Oracle Text CONTEXT indexes are not synchronous. That is, updates to the table are not immediately reflected in the index. Instead, the index must be synchronized. We can do this manually with a call to CTX_DLL.SYNC_INDEX, or we can arrange to have it done automatically for us.

Let's show the manual method first:. We'll insert a new run then search for it. We won't find it because the index is not up-to-date. Then we'll call SYNC_INDEX, giving it the name of the index, and search again:

SQL> insert into quickstart values (4, 'GB', 'Michael Smith');
1 row created.

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'michael', 99) > 0 order by score(99) desc;
no rows selected

SQL> execute ctx_ddl.sync_index('full_name_index')
PL/SQL procedure successfully completed.

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'michael', 99) > 0 order by score(99) desc;

 SCORE(99)         ID FULL_NAME
---------- ---------- ----------------------------------------
         5          4 Michael Smith

If we don't want to bother with manual SYNC calls,  we can use the parameter string "SYNC (ON COMMIT)" when we create the index. That means that immediately after updates are committed to the table, a SYNC operation will be run to get the index into step with the table.

SQL> drop index full_name_index;
Index dropped.

SQL> create index full_name_index on quickstart( full_name ) indextype is ctxsys.context parameters ('sync (on commit)');
Index created.

SQL> insert into quickstart values (5, 'US', 'Scott Peters');
1 row created

We can search for that, but we won't find it since we haven't yet committed - so the SYNC has not run: 

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'scott', 99) > 0 order by score(99) desc;
no rows selected

Now we can commit...

SQL> commit;
Commit complete.

And this time it will find it:

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'scott', 99) > 0 order by score(99) desc;

 SCORE(99)         ID FULL_NAME
---------- ---------- ----------------------------------------
         5          5 Scott Peters

SYNC(ON COMMIT) is fine if you don't make too many updates to your table, but is not ideal if you have a very large table and lots of changes take place on it.

Because of the way CONTEXT indexes are built, frequent changes will cause fragmentation of the index, decreasing performance over time. So a common way is to arrange for indexes to be synchronized at certain fixed intervals.  We can do this manually, using the database scheduler, or we can have it done automatically with a "SYNC (EVERY ... )" string in the parameters clause.  The time clause in that uses scheduler syntax, and can be a little convoluted. Every 1 minute can be represented in days as 1/24/60", so we can try the following:

SQL> drop index full_name_index;
Index dropped.

SQL> create index full_name_index on quickstart( full_name ) indextype is ctxsys.context parameters ('sync ( every SYSDATE+1/24/60 )');
Index created.

SQL> insert into quickstart values (6, 'US', 'Paul Peters');
1 row created.

SQL> commit;
Commit complete.

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'paul', 99) > 0 order by score(99) desc;

no rows selected

So we've committed, but the regular SYNC hasn't kicked off yet. We can wait a minute and try again, and this time we'll find it:

SQL> select score(99), id, full_name from quickstart where contains ( full_name, 'paul', 99) > 0 order by score(99) desc;

 SCORE(99)         ID FULL_NAME
---------- ---------- ----------------------------------------
         5          6 Paul Peters

Back in Part 1 we said that it was necessary to grant the "CREATE JOB" privilege to our user - this is why.  If the user doesn't have CREATE JOB privilege, then we will get an "insufficient privileges" error if we try to create the index with "SYNC (EVERY ...)".

Finally, we should talk about index optimization. CONTEXT indexes, over time, get less efficient as they get updated.  This inefficiency takes two forms:

  1. The index gets filled with garbage - it contains references to deleted or updated documents which are no longer current
  2. The "postings lists" get fragmented. The pointers to which documents contain each word get broken up into small chunks instead of the idea long strings.

We can fix this by running index optimization.  As with the manual call to sync indexes, this is done with a call to the PL/SQL package CTX_DDL:

SQL> execute ctx_ddl.optimize_index('full_name_index', 'FULL')
PL/SQL procedure successfully completed. 

The "FULL" parameter there is the mode of optimization. You can choose from the following:

  • FAST: Only posting fragmentation is fixed - no garbage collection occurs
  • FULL: Both fragmentation and garbage collection is dealt with
  • REBUILD: The entire index is copied to a fresh table (requires more disk space but produces the best results).

The more frequently you run SYNC, the more fragmented your index will become, and the more often you will need to optimize.

A typical "maintenance regime" at many customers is to run SYNC every five minutes, then run OPTIMIZE in FULL mode nightly, and OPTIMIZE in REBUILD mode once a week. 

The next topic is for advanced users: Getting Started Part 4 - Advanced Index Customization

Monday Jul 28, 2014

Getting started Part 2 - More queries

[Read More]

Thursday Jul 24, 2014

Getting started with Oracle Text

My previous posts have mostly been aimed at experienced Oracle Text users. But what if you (or your colleagues) have never used it before?  What are the basics to get started? I hope this post will help some of those people.

Because I'm old-school (or maybe just old), my examples will be listed as though you're entering them in command-line SQL*Plus. But they will work just as well if you use them in SQL Developer.

First of all, any user of Oracle Text needs a special role CTXAPP.   Let's create a new user DEMO with that role, as well as the common CONNECT and RESOURCE roles - and also CREATE JOB which we'll need later in this series of tutorials.

create user demo identified by demo default tablespace users temporary tablespace temp quota unlimited on users;

grant connect, resource, ctxapp, create job to demo;

Now if we're going to call Oracle Text functionality from within PL/SQL procedures, we need some extra grants (since grants to roles don't work in packages).  See Step 3 in the Text Developer's Guide. We won't cover that here.

Now we'll assume the user has an application which queries a table called QUICKSTART.  For the demo, we should login as the DEMO user we just created, and run:

create table quickstart (id number primary key, country varchar2(2), full_name varchar2(40));

 We'll insert some records into that:

insert into quickstart values (1, 'US', 'John Doe');

insert into quickstart values (2, 'GB', 'John Smith');

insert into quickstart values (3, 'NZ', 'Peter Smith-Smith');

Now a simple query against that table might look like

select * from quickstart where country = 'GB';

Now let's say we want to search for anyone with "Smith" in their name.  We could do this by simply extending the query to:

select * from quickstart where country = 'GB' and upper(full_name) like '%SMITH%';

but there are a few problems with this:

  • There is no index. The kernel must read every full_name field and scan it. Not a problem here, but certainly a problem if there are millions of records.
  • The query would match SMITHSON and BLACKSMITH. Not a problem if that's what we want, but tricky otherwise.

We can create a word-based index on the FULL_NAME column using the following syntax:

create index full_name_index on quickstart( full_name ) indextype is ctxsys.context;

 Note this is like any other "create index" statement, with the addition of the phrase "indextype is ctxsys.context", telling the kernel to create a specialized CONTEXT index and allow the use of query operators associated with that indextype.

For example we can do:

 select * from quickstart where country = 'GB' and contains ( full_name, 'smith') > 0;

A few things to note:

  • The CONTAINS operator can only be used if an index is present on the column specified in the first argument
  • The search string doesn't need any wildcards either side.  We are looking for the whole word 'smith' (if we wanted to match 'smithson' as well, then we could use a wildcard: 'smith%').
  • SQL doesn't have boolean functions, so CONTAINS returns zero for a non-match, and greater-than-zero for a match.  The actual value returned is the 'score' (see later) but there are very rarely any circumstances where you do anything other than testing where the return value is greater than zero.

Scoring: Most SQL queries just select between rows that match a criterion, and those that don't. There is rarely the concept of a good or better match. But if we're searching a lot of technical papers for the word "Exadata", then a document which has many occurrences of this term is likely to be a better match for our search than one where there is only a single match. Therefore CONTEXT indexes return a score for each search.  As mentioned, that is returned by the CONTAINS clause, but is rarely useful there.  It's usually better to use it in the SELECT clause, and perhaps in an ORDER BY clause as well.  A score is associated with a particular CONTAINS clause, of which a query may have several, so we use a number as the third argument in the CONTAINS function, and SCORE operator takes the same number.  It doesn't matter what that number is so long as it's the same in both cases.  We'll use 99 in this example:

select score(99), id, full_name from quickstart where contains ( full_name, 'smith', 99) > 0 order by score(99) desc;

This produces the output:

 SCORE(99)         ID FULL_NAME
---------- ---------- ------------------------------
         7          3 Peter Smith-Smith
         4          2 John Smith

Note that the first item here scores higher than the second, because the search term smith appears twice.  Note there is no "absolute" meaning to those scores - you can't say  the first is a "7% match" or anything like that - all you can say is that a record with a higher score is more relevant than one with a lower score.

Scoring is quite a complex topic, which we'll cover in more detail later.

Move ahead to Part 2 - More queries

Monday Jan 06, 2014

Datastore triggers in 12c

If you want to create an Oracle Text CONTEXT index on multiple columns, you would normally use a MULTI_COLUMN_DATASTORE or USER_DATASTORE to bring in the various columns for indexing.

This works well, but the kernel doesn't know which columns you've used in the index.  As far as the kernel is concerned, you have created the index on a specific column, and it will only update the index when the contents of that column change.

Hence it is common to create a trigger something like the following:

CREATE OR REPLACE TRIGGER customer_update_trg
BEFORE UPDATE OF lastname ON customers
  :new.firstname := :new.firstname;

In this case, the index was created on the firtname column, but the multi_column_datastore fetches in both lastname and firstname.  The trigger ensures that whenever lastname is changed, we force an update of firstname in order to reindex the row.

However, things have changed in 12c. As a fix for bug 14155722, triggers check the "before" and "after" state of any referenced column values, and if the column has not changed, then no index updates take place.  So the trigger above won't cause an index update if only lastname has changed.

The original behavior can be restored by an alter session or alter system operation, as follows:

alter session set "_fix_control"='14155722:OFF'


alter system set "_fix_control"='14155722:OFF';

If done as an alter session, it must be done before the update operation which invokes the trigger, and will need ALTER SESSION privilege.  If done as alter system it will apply globally, but must be run by a DBA user.  It can also be put into the database initialization file, either by adding "scope=spfile" to the alter system command, or if a pfile is in use by adding it explicitly there.

Wednesday Sep 25, 2013

High Volume Indexing with Oracle Text

My presentation from OpenWorld San Francisco 2013 is available here: High Volume Indexing with Oracle Text.

The Hands On Lab workbook is here: Text Analytics with Oracle Text.  Note this doesn't include scripts, or any data for the Classification section.

Friday Mar 22, 2013

Using Entity Extraction

Entity Extraction using Oracle Text was introduced with the release of the Oracle Database. It's not a particularly well-known feature, so I thought I'd make some notes on how to use it.

Please note that Entity Extraction uses some third party technology "under the covers" which is deliberately disabled in If you run these examples in (or when it's out) you will receive a "feature not generally available" message. The functionality can be re-enabled, but only by contacting Oracle Support so they can run some checks on whether your applications will be affected by the replacement of this third party technology in the next major release.

Entity extraction is the capability to identify, and extract, named entities within a text.

Entities are mainly nouns and noun phrases. They include names, places, times, coded strings (such as phone numbers and zip codes), percentages, monetary amounts and many others.

The ctx_entity package implements entity extraction by means of a built-in dictionary and set of rules for English text. The capabilities can be extended for English, or for other languages, by means of user-provided add-on dictionaries and rule sets.

Let's look at a very simple, basic example.  

Let's assume we have a clob containing the text:

New York, United States of America <p>
The Dow Jones Industrial Average climbed by 5% yesterday on news of a new software release from database giant Oracle Corporation.'

We will use ctx_entity.extract to find all the entities in the CLOB value. (For now, we won't worry about how the text got into the clob or how we provide the output clob - that's all in the examples if you want to look).

Entity extraction requires a new type of policy - an "extract policy" - which allows you to specify options. For now, we will create a default policy, thus:

ctx_entity.create_extract_policy( 'mypolicy' );

We can then call extract to do the work. It needs four arguments: the policy name, the document to process, the language, and the output clob (which must have been initialized, for example by calling dbms_lob.createtemporary).

ctx_entity.extract( 'mypolicy', mydoc, 'ENGLISH', outclob )

outclob contains the XML identifying extracted entities.  If we display the contents (preferably by selecting it as an xmltype so it gets formatted nicely) we will see:

  <entity id="0" offset="0" length="8" source="SuppliedDictionary">
    <text>New York</text>
  <entity id="1" offset="150" length="18" source="SuppliedRule">
    <text>Oracle Corporation</text>
  <entity id="2" offset="10" length="24" source="SuppliedDictionary">
    <text>United States of America</text>
  <entity id="3" offset="83" length="2" source="SuppliedRule">
  <entity id="4" offset="113" length="8" source="SuppliedDictionary">
  <entity id="5" offset="0" length="8" source="SuppliedDictionary">
    <text>New York</text>

That's fine if we're going to process it with an XML-aware program. However, if we want it in a more "SQL friendly" view, we can use XML DB functions to convert it:

select xtab.offset, xtab.text, xtab.type, xtab.source
from xmltable( '/entities/entity'
PASSING xmltype(outclob)
    offset number       PATH '@offset',
    lngth number        PATH '@length',
    text   varchar2(50) PATH 'text/text()',
    type   varchar2(50) PATH 'type/text()',
    source varchar2(50) PATH '@source'
) as xtab order by offset;

which produces as output:

    OFFSET TEXT                      TYPE                 SOURCE
---------- ------------------------- -------------------- --------------------
         0 New York                  city                 SuppliedDictionary
         0 New York                  state                SuppliedDictionary
        10 United States of America  country              SuppliedDictionary
        83 5%                        percent              SuppliedRule
       113 software                  product              SuppliedDictionary
       150 Oracle Corporation        company              SuppliedRule

Hopefully that's fairly self-explanatory.  Each entity found is listed with it's offset (starting point within the text), length, text, entity type, and source (how it was found).  The entity "New York" is identified as both a City and a State, so it appears twice.

If we don't want all the different types of entity, we can select which types to fetch. We do this by adding a fourth argument to the "extract" procedure, with a comma-separated list of entity types.  For example:

ctx_entity.extract( 'mypolicy', mydoc, 'ENGLISH', outclob, 'city, country' )

That would give us the XML

  <entity id="0" offset="0" length="8" source="SuppliedDictionary">
    <text>New York</text>
  <entity id="2" offset="10" length="24" source="SuppliedDictionary">
    <text>United States of America</text>

Next let's look at creating a new entity type using a user-defined rule. Rules are defined using a regular-expression-based syntax.
The rule is added to an extraction policy, and will then apply whenever that policy is used

We will create a rule to identify increases, for example in a stock index.

There are many ways to express an increase - we're hoping to match any of the following expressions:
  climbed by 5%
  increased by over 30 percent
  jumped 5.5%

So we'll create a regular expression which matches any of those, and create a new type of entity. User defined entities must start with the letter "x", so we'll call our entity "xPositiveGain":

ctx_entity.add_extract_rule( 'mypolicy', 1,
    '<rule>'                                                          ||
      '<expression>'                                                  ||
         '((climbed|gained|jumped|increasing|increased|rallied)'      ||
         '( (by|over|nearly|more than))* \d+(\.\d+)?( percent|%))'    ||
      '</expression>'                                                 ||
      '<type refid="1">xPositiveGain</type>'                          ||

Note the "refid" in there. This tells us which part of the regular expression to actually match, by referencing a pair of parentheses within it. In our case, we want the entire expression, so that's the outermost (and first occurring) parentheses, so that's refid=1.

This time, it is necessary to compile the policy before we can use it:


Then we can use it as before

  ctx_entity.extract('mypolicy', mydoc, null, myresults)

The (abbreviated) output of that would be

  <entity id="6" offset="72" length="18" source="UserRule" ruleid="1">
    <text>climbed by over 5%</text>

Finally, we're going to add another user defined entity, but this time as using a dictionary.  We want to recognize "Dow Jones Industrial Average" as an entity of type xIndex.  While we're at it we'll add "S&P 500" as well. To do that, we create an XML file containing the following:

      <value>dow jones industrial average</value>
      <value>S&amp;P 500</value>

Case is not significant in this file, but note how the "&" in "S&P" must be specified as the XML entity "&amp;" - otherwise the XML would not be valid.

This XML file is loaded into the system using the CTXLOAD utility.  If the file was called "dict.load", we would use the command:

ctxload -user username/password -extract -name mypolicy -file dict.load

Again we need to compile the policy using ctx_entity.compile. Then when we run ctx_entity.extract we will see in the output:

  <entity id="6" offset="72" length="18" source="UserRule" ruleid="1">
    <text>climbed by over 5%</text>
  <entity id="7" offset="43" length="28" source="UserDictionary">
    <text>Dow Jones Industrial Average</text>

Here's some example code so you can try it yourself.  It uses SQL*Plus bind variables, which don't work too well in SQL Developer - if you really don't like using SQL*Plus, you'll need to convert it - perhaps put the whole thing into a PL/SQL procedure and use PL/SQL variables instead of SQL*Plus variables.

Download link (right-click and "Save As" on most systems): entity_extraction.sql

Friday Mar 15, 2013

A staging table for Oracle Text index updates

Oracle Text CONTEXT indexes differ from standard BTREE indexes in that they are not synchronous.  That is, the index is not updated at the same time that the data changes. It is necessary to invoke an index sync in some manner before any updates (or inserts) to the table are reflected in the index.

The simplest way to do this is to specify SYNC (ON COMMIT) in the parameters clause of the index creation. That means that as soon as changes are committed, the index is sync'd.

 (This still isn't quite synchronous. The indexing doesn't start until the commit, so although changes to the base table will be visible to the user who made them before doing a commit, that user will NOT see changes reflected in the index. Additionally, there will be a short period between the end of the commit and the end of the sync when other userswill be able to see the table changes but not the index changes).

 SYNC(ON COMMIT) is popular, but it's not necessarily the best option.  This is because frequent syncs cause fragmentation of the index. New data is written in small chunks to the end of the $I postings list table, and the more and smaller these are, the worse query performance will be until the index is optimized.

So there is usually a trade-off between update frequency, and the fragmentation of an index. Users would like to have their updates available for searching immediately by means of SYNC(ON COMMIT), but DBAs would prefer to have SYNCs spaced widely apart to avoid fragmentation of the index.

One way around this is to create a staging table to hold recent updates to the main table.  The staging table will be a copy of the main table's layout, with the addition of a timestamp column.

Indexes are created on both tables (main and stage).  These indexes are identical, except the main index has manual sync, and the staging table has SYNC(ON COMMIT).

We then create triggers which copy inserts and updates from the main table to the stage table, where they are indexed near-instantaneously.  Because the stage table is kept small, we needn't worry about fragmentation.  It's probably advisable to arrange things such that the stage table is "pinned" in the SGA buffer cache, but if it's used regularly, it will probably stay in memory anyway.

Queries which would previously have run against the main table are modified so that they do a UNION between the main table and the staging table. That way, we get all the unmodified rows from the main table, and the modified rows from the stage table.  If you're thinking "won't we get the modified rows in their old form from the base table?" then remember that DELETEs are actioned immediately in a CONTEXT index, and that updates are implemented internally as DELETE followed by INSERT).

Then we just need to put a wrapper around the SYNC for the main table which performs the actual SYNC then deletes all the rows from the staging table which were inserted before the start of the SYNC (hence the need for a timestamp column).  If we choose, we can also optimize the stage table index at this point to.

I've written some sample code to demonstrate this on a simple table. It is intended to be run in SQL*Plus but will probably work in SQL Developer with little or no modification.  Be aware that it drops and recreates a user called TESTUSER.  You will need to edit the CONNECT line unless your SYSTEM password is "manager".

Download link (right click and save): staging_index.sql or view in browser: staging_index.sql.txt

Friday Feb 01, 2013

Partition Maintenance and Oracle Text Indexes

Oracle Text supports "locally partitioned indexes" on partitioned tables.

That means that if you have a RANGE partitioned base table, you can use the LOCAL keyword when you create your index, and Oracle Text will automatically produce a partitioned index, with one index partition for each partition of the base table.

A partitioned index is very much like having a separate Oracle Text index on each partition. There is a $I table, a $K table, a $R table, etc, for each partition. The naming convention for the index is changed slightly - the $I table for an unpartitioned table is DR$index_name$I, for a partitioned table it is DR#index_name0001$I, DR#index_name0002$I, etc (the eagle-eyed will have spotted an apparent limit of 9,999 partitions, but in fact this is lifted in 11.2, as it moves to an alphunumeric naming convention after 9999).

Note that currently only RANGE partitioned tables are supported with local Text indexes.

An important feature of any partitioned table is the ability to split and merge partitions. Splitting means dividing one existing partition into two new partitions, and merging means combining two (or sometimes more) partitions into a single one.

Normally, any LOCAL indexes on partitions which have been split or merged become invalid - they must be rebuilt before they are useable. However, if you specify UPDATE GLOBAL INDEXES, the local indexes will be automatically rebuilt as part of the split or merge:
ALTER TABLE my_table
That's fine - but if we've got lots of data in there, it's going to take quite some time to rebuild those indexes, and our application is going to be unavailable for all that time.

Is there a way we can do a split or merge without having to take the application offline while we rebuild the indexes? Yes, there is. We can make use of a feature known as Partition Exchange, which allows us to swap a complete table, including its ready-built indexes, in place of an existing partition in a partitioned table.  The syntax looks like:
ALTER TABLE my_table
The WITHOUT VALIDATION is optional, but makes it a lot quicker. It avoids the need for the system to check every partition key value to make sure the row is going into the correct partition. Obviously it needs to be used with care.

So now we have a solution.  When we do a merge, we first copy the data from the partitions to be merged into a temporary table. We create an index on the temporary table using all the same index options as for our main table. When that index is complete, we do the ALTER TABLE ... MERGE PARTITION as above, but without the "UPDATE GLOBAL INDEXES" clause (thus leaving our indexes invalid for a short period) and then we do the ALTER TABLE ... EXCHANGE PARTITION to swap in the partition complete with its freshly built index in place of the merged partition with its invalid index.

We can do something similar with SPLIT PARTITION. This time, though, we will need two temporary tables.  We copy the data out of the first part of the partition to be split into one temporary table, and the data from the second part of the partition to be split into the other temporary table.  We create indexes on both of those temporary tables (we can build the two indexes simultaneously if we have enough resources, or sequentially if not.  We then split the partition, without using UPDATE GLOBAL INDEXES:
ALTER TABLE my_table 
and then we do two partition exchanges
ALTER TABLE my_table
and again for the second
ALTER TABLE my_table

... and that's it. Or is it? What about changes which have been made to the base table since we copied the data out into temporary tables?

Well the simple answer to that is to make sure we don't do any changes to the table. We could even make it read-only:


That solves the problem, but it's rather heavy-handed. After all, the reason we're doing all this is that it can take a long time to build text indexes, and we would prefer not to have to prevent updates to the table.  So instead we need to institute some kind of "change management".  There are doubtless several ways to achieve this, but I've done it by creating triggers which monitor any updates or inserts on the base table, and copy them to a temporary "staging" table. These transactions can then be copied back to the main table after the partition split or merge is complete, and the index sync'd in the normal way.

I've provided a pair of complete examples which illustrate this process. They do a split and a merge respectively on a simple partitioned table, creating indexes on temporary tables as necessary and exchanging those tables in place of partitions in the main table.  They also create the necessary triggers to monitor updates, and roll these back into the main table on completion.

One note about the example: I've partitioned the tables on a the "price" column.  While this seemed a good idea in the context of a simple, understandable "e-commerce" type model, it's actually a pretty bad idea, since you can't change the value of a partition key if it would cause the row to move partitions.  So if we have a partition boundary at 50, and an item which costs 49, then we can't increas its price to 51.


The "download links" are for saving to a file - on most browsers you can right click and "Save Target As" or "Save Link As". 

If you want to view the code in a browser just click on "open in browser" (the file will have an extra .txt suffix if you then save it from the browser page).

split_partitions.pls - download link / open in browser

merge_partitions.pls - download link / open in browser

Wednesday Jul 11, 2012

Oracle Text query parser

Oracle Text provides a rich query syntax which enables powerful text searches.

However, this syntax isn't intended for use by inexperienced end-users.  If you provide a simple search box in your application, you probably want users to be able to type "Google-like" searches into the box, and have your application convert that into something that Oracle Text understands.

For example if your user types "windows nt networking" then you probably want to convert this into something like
"windows ACCUM nt ACCUM networking".  But beware - "NT" is a reserved word, and needs to be escaped.  So let's escape all words:
"{windows} ACCUM {nt} ACCUM {networking}".  That's fine - until you start introducing wild cards. Then you must escape only non-wildcarded searches:
"win% ACCUM {nt} ACCUM {networking}".  There are quite a few other "gotchas" that you might encounter along the way.

Then there's the issue of scoring.  Given a query for "oracle text query syntax", it would be nice if we could score a full phrase match higher than a hit where all four words are present but not in a phrase.  And then perhaps lower than that would be a document where three of the four terms are present.  Progressive relaxation helps you with this, but you need to code the "progression" yourself in most cases.

To help with this, I've developed a query parser which will take queries in Google-like syntax, and convert them into Oracle Text queries. It's designed to be as flexible as possible, and will generate either simple queries or progressive relaxation queries. The input string will typically just be a string of words, such as "oracle text query syntax" but the grammar does allow for more complex expressions:

  word : score will be improved if word exists
  +word : word must exist
  -word : word CANNOT exist
  "phrase words" : words treated as phrase (may be preceded by + or -)
  field:(expression) : find expression (which allows +,- and phrase as above) within "field".

So for example if I searched for
  +"oracle text" query +syntax -ctxcat
Then the results would have to contain the phrase "oracle text" and the word syntax. Any documents mentioning ctxcat would be excluded from the results.

All the instructions are in the top of the file (see "Downloads" at the bottom of this blog entry).  Please download the file, read the instructions, then try it out by running "parser.pls" in either SQL*Plus or SQL Developer.

I am also uploading a test file "test.sql". You can run this and/or modify it to run your own tests or run against your own text index. test.sql is designed to be run from SQL*Plus and may not produce useful output in SQL Developer (or it may, I haven't tried it).

I'm putting the code up here for testing and comments. I don't consider it "production ready" at this point, but would welcome feedback.  I'm particularly interested in comments such as
  • "The instructions are unclear - I couldn't figure out how to do XXX"
  • "It didn't work in my environment" (please provide as many details as possible)
  • "We can't use it in our application" (why not?)
  • "It needs to support XXX feature"
  • "It produced an invalid query output when I fed in XXXX"


The "download links" are for saving to a file - on most browsers you can right click and "Save Target As" or "Save Link As". 

If you want to view the code in a browser just click on "open in browser" (the file will have an extra .txt suffix if you then save it from the browser page).

parser.pls - main program file download link / open in browser

test.sql - for testing procedure download link / open in browser

Wednesday May 16, 2012

Oracle Text and accented characters

There are a number of lexer attributes in Oracle Text that affect the way accented characters are dealt with. We're going to look at each of these and consider the effects they have, alone and together.

Before we look into them, a quick aside on typing accented characters if you don't have them natively on your keyboard. I'm not going to go too deep into character sets and code pages, but if you're on a European Windows machine, you're probably using the WIN1252 character set.

If I want to type one of the accented characters in that set, in most Windows applications I can hold down the ALT key, then enter the four-digit decimal code for that character on my numeric keypad. So to enter a lower-case o-umlaut character ("o" with two dots above it: "ö"), I would hold down ALT and type 0246 on the numeric keypad.


Let's start with perhaps the simplest - the BASE_LETTER attribute. If I set this to "true" then characters with accents in the text will effectively be indexed as their base form.

So for example if I index the German word "schön" (meaning beautiful in English) then the actual indexed token will be "schon". That means if the user searches for " schön" or "schon" then the word will be found (since the query term is processed in the same way).

That makes life much easier for people who don't have the oh-umlaut character on their keyboard, and don't know the ALT trick mentioned above (or don't want to have to look up the necessary character code).

So that's simple - might as well set that option on always, right? Well, not necessarily. Because in German, the word "schon" actually has a quite different meaning to " schön", and German users wouldn't want to find the unaccented word if they specifically meant to look for the accented word.

So the simple rule of thumb is this: If you think that most of your users will be searching for words for which they have the correct keyboard (for example German users searching German text) then it's best to set BASE_LETTER to false. But if you think users might want to search foreign words for which they do not have the correct keyboard (for example English, or multi-national users searching German or mixed text) then it's best to set BASE_LETTER to TRUE.


Now it starts to get a little more complex. Some languages - notably German again - allow you to avoid the use of accented characters by spelling the words differently. In German, an accented character is replaced by the un-accented form of the same letter, followed by an "e". So "shön" could equally validly be spelled "shoen", and every German user would recognise it as the same word. Equally "Muenchen" (the city English speakers call Munich) would be recognized as the same as "München".

So that we can treat these alternate spelling forms as equivalent Oracle Text has the ALTERNATE_SPELLING attribute. When set to "german", Oracle Text will look for the "alternate form" of "vowel + e" and index both that and the accented form of the word. When processing a query, it will equally translate "oe" into "ö", etc, thus ensuring that the word can always be found, regardless of which of the alternate forms is searched for, and which is in the indexed text.

Aside: anyone following closely might ask "why does it index both the alternate form "shoen" and the accented form "shön"? Surely it would be sufficient to just index the "shön" form? Well, mostly it would. But what happens if the word in question was actually an English word in the middle of the German text, such as "poet"? OK, it would be indexed as "pöt" and anybody seaching for "poet" would still find it (since the transformation is applied to the searchterm as well). But what if they used a wildcard and searched for "po%"? They would still expect to find the term, but if only the accented form was indexed, they wouldn't. Hence the reason we index both forms, just in case.


OK, so we want ALTERNATE_SPELLING set to "german" for our German text. But we also know that people with non-German keyboards are often going to search it. So we BASE_LETTER on as well. What happens now?

If the indexer comes across "schön", ALTERNATE_SPELLING would normally index that without any change. But BASE_LETTER forces it to the unaccented form, and "schon" is actually indexed. If the indexer comes across "shoen", then ALTERNATE_SPELLING decides it should be indexed as both "schoen" and "schön". But before the tokens are written to the index, BASE_LETTER is applied, so the tokens "shoen" and "shon" are indexed.

That all works fine, and we can find either term by searching for "shon", "shön" or "shoen". Great!

But (there's always a but) what happens if we index the French word "Rouède" (the name of a town near the Spanish border)? "uè" is not a candidate for ALTERNATE_SPELLING, so it is left alone. Then BASE_LETTER is applied, and the word "Rouede" is written to the index. If the user searches for "Rouède" then the query-time processing works the same, the search is converted to "Rouede" and the query works fine. However, if the user searches for the base letter form "Rouede", things don't go so well. This time ALTERNATE_SPELLING does get applied to the query term (since the query processor has no way of knowing that the "e" character should be accented) and the searchterm is converted to "Roüde". BASE_LETTER is then applied, and it looks for "Roude" in the index. But the indexed term is "Rouede", so nothing is found.


To solve this problem, the OVERRIDE_BASE_LETTER attribute was introduced.

If you set OVERRIDE_BASE_LETTER to "true", then ALTERNATE_SPELLING will "mask" BASE_LETTER. That means that if we meet accented characters in the text which have a alternate form (such as "ö"), we will index them in their original, accented form and also in their alternate form. If we meet them in their alternate form (eg Muenchen) we will index ONLY the alternate form and not transform them.  Accented characters which do not have an alternate form (such as "è") have BASE_LETTER processing applied to them to transform them to their equivalent unaccented character.  Then at query time, we apply only ALTERNATE_SPELLING to any appropriate accented search terms, and BASE_LETTER to all others.This has the positive effect that our previous example "rouède" can be found, if searched for with or without the accent on the "e" character.

It does have the negative effect that base letter searches no longer work on German words - we can't search for "shon" anymore, only "shön" or "shoen" will work. So OVERRIDE_BASE_LETTER makes sense if we want to perform ALTERNATE_SPELLING on German (or other specified language) words, and BASE_LETTER on all other languages.

Appendix: Test Script

This is the script I used to test the effects of the various options. To avoid any issues with character set translation, I used the UNISTR() function to create Unicode characters for my accented characters. Note the German words are prefixed by two-letter codes "wa" - with accent, "na" - no accent and "af" alternate form. That allowed me to distinguish in the index between the index terms derived from "schön" and those derived from "schon".

ctx_ddl.drop_preference ( 'my_lexer');

ctx_ddl.create_preference ( 'my_lexer', 'BASIC_LEXER' );
ctx_ddl.set_attribute( 'my_lexer', 'BASE_LETTER', 'true' );
ctx_ddl.set_attribute( 'my_lexer', 'OVERRIDE_BASE_LETTER', 'true');
ctx_ddl.set_attribute( 'my_lexer', 'ALTERNATE_SPELLING','german' );

drop table tt;
create table tt(a1 number primary key,text varchar2(45));

-- town name "Rouède", accent on the e
insert into tt values (1,'rou'||unistr('\00E8')||'de');

-- shön with accent (wa)
insert into tt values (2,'wash'||unistr('\00F6')||'n');

-- shon no accent (na)
insert into tt values (3,'nashon');

-- muenchen alternate form (af)
insert into tt values (4,'afmuenchen');


select * from tt;

create index tta on tt(text) indextype is ctxsys.context
parameters ( 'lexer my_lexer' );

set feedback 2

select token_text, token_type from dr$tta$i;

PROMPT searching for the base letter form, without accent on the first e
select * from tt where contains(text,'Rouede')>0;

PROMPT and with the accent
select * from tt where contains(text,'Rou'||unistr('\00E8')||'de') > 0;

--select * from tt where contains(text,'m'||unistr('\00FC')||'nchen')>0;
--select * from tt where contains(text,'muenchen') > 0;

set echo on

--select * from tt where contains(text,'nashoen') > 0;
--select * from tt where contains(text,'nashon') > 0;
--select * from tt where contains(text,'na'||unistr('\00F6')||'n') > 0;

select * from tt where contains(text,'washon') > 0;
select * from tt where contains(text,'washoen') > 0;
select * from tt where contains(text,'wa'||unistr('\00F6')||'n') > 0;

set echo off

-- The following section shows how to see how the query has been transformed - it shows
-- the actual words looked for in the index.

drop table test_explain;
create table test_explain(
explain_id varchar2(30),
id number,
parent_id number,
operation varchar2(30),
options varchar2(30),
object_name varchar2(64),
position number,
cardinality number);

index_name => 'tta',
text_query => 'wasch'||unistr('\00F6')||'n',
explain_table => 'test_explain',
sharelevel => 0,
explain_id => 'Test');

col explain_id for a10
col id for 99
col parent_id for 99
col operation for a10
col options for a10
col object_name for a20
col position for 99

select explain_id, id, parent_id, operation, options, object_name, position
from test_explain order by id;

Wednesday Apr 18, 2012

Oracle Text on Solid State Disks

We've known for some time that putting Oracle Text indexes onto flash-memory-based solid state disks (SSDs) gives some big advantages.

A while back, we tried benchmarking Oracle Text on Fusion I/O devices, which showed huge improvements. But those PCI-based devices are expensive, whereas consumer level SSDs are getting cheaper and cheaper. I decided to compare some Oracle Text queries on an inexpensive SSD against the same queries on a rotational disk.

Test Hardware:

Dell D6410 laptop
2.4GHz Intel i5 CPU (2 cores, 4 threads)
Mushkin Chronos 120GB SSD
Seagate 750GB conventional disk (5400 rpm)

How I tested

First I created two Oracle users, "flash" and "nonflash". Then I created two tablespaces of 20GB each, one on the SSD and one on the conventional disk. I set the default tablespace for each user to the appropriate one.

Next I loaded five million rows of random text, totalling about 10GB.

For this I used a oci C program "randomtext_oci.c". This creates random length texts up to 4K in length, composed of random word lengths and random characters. The words are composed of a fairly small set of characters - the consonants BCDFG alternating with the vowels AEIO. Although this is not really representative of real text, it's better than completely random characters. In our 10GB of text, there are around 210,000 unique words, which is comparable to what you'd expect on a similarly-sized collection of clean text (by "clean", I mean very few spelling mistakes or made-up words).

This text was then copied from one user to the other (using CREATE TABLE mydocs AS SELECT * FROM flash.mydocs UNRECOVERABLE)
and each table indexed with a simple Oracle Text index:

create index mydocs_index on mydocs(text) indextype is ctxsys.context
parameters ('memory 500000000')

The index took around 40 minutes to create. It was a little quicker on the SSD, but not much.

I then ran several queries to check the performance with different numbers of hits returned.

Each query took the form "SELECT COUNT(*) FROM mydocs WHERE CONTAINS (text, '<searchterms>') > 0;

Using COUNT(*) means that we're testing only the cost of fetching rowids from the text index - if we fetched other data from the table we'd be measuring the cost of fetching base table blocks as well.

Before running each query, I rebooted the machine (thankfully quite quick as the OS is on SSD!) then pre-loaded the $R table using the LoadAllDollarR procedure as documented here. The $R table generally gets fully cached in memory within a short time, so we are mainly interested in the time needed to fetch the $I table info.  I also ran a query which fetched zero results, to make sure the Oracle Text data dictionary was loaded into cache.

Here are the queries, and the results I found:

 Query  No. of Hits
 Conventional Disk
 SSD  Difference
 bab  1,283,890  1.14 sec 0.53 sec
 2.2 x
 bad OR bac
  2,149,822  2.73 sec 0.90 sec
 3.0 x
 ba  3,478,811  5.44 sec 1.23 sec  4.4 x
 b  4,609,044  8.26 sec 1.95 sec  4.2 x

We can see this graphically (number of hits on the horizontal axis, number of seconds on the vertical axis)

Performance graph


Because Oracle Text queries require lots of random I/O to fetch "posting lists" from the index, we see big improvements from using solid state disks.  Although the advantages would be less marked if we compared against faster rotational disks (such as 15K rpm SAS disks), we would still most likely see a significant improvement.  And the SSD used here was a very inexpensive model (at around US $1 per GB), so we might expect to see even better improvements using a more advanced professional device, or by using a PCI-based solution rather than a SATA device.

Tuesday Apr 10, 2012

Indexing data from multiple tables with Oracle Text

It's well known that Oracle Text indexes perform best when all the data to be indexed is combined into a single index.

The query

  select * from mytable
  where contains (title, 'dog') > 0
     or contains (body, 'cat') > 0
will tend to perform much worse than
  select * from mytable 
  where contains (text, 'dog WITHIN title OR cat WITHIN body') > 0
For this reason, Oracle Text provides the MULTI_COLUMN_DATASTORE which will combine data from multiple columns into a single index. Effectively, it constructs a "virtual document" at indexing time, which might look something like:

<title>the big dog</title>
<body>the ginger cat smiles</body>

This virtual document can be indexed using either AUTO_SECTION_GROUP, or by explicitly defining sections for title and body, allowing the query as expressed above. Note that we've used a column called "text" - this might have been a dummy column added to the table simply to allow us to create an index on it - or we could created the index on either of the "real" columns - title or body.

It should be noted that MULTI_COLUMN_DATASTORE doesn't automatically handle updates to columns used by it - if you create the index on the column text, but specify that columns title and body are to be indexed, you will need to arrange triggers such that the text column is updated whenever title or body are altered.

That works fine for single tables. But what if we actually want to combine data from multiple tables?

In that case there are two approaches which work well:

  1. Create a real table which contains a summary of the information, and create the index on that using the MULTI_COLUMN_DATASTORE. This is simple, and effective, but it does use a lot of disk space as the information to be indexed has to be duplicated.
  2. Create our own "virtual" documents using the USER_DATASTORE. The user datastore allows us to specify a PL/SQL procedure which will be used to fetch the data to be indexed, returned in a CLOB, or occasionally in a BLOB or VARCHAR2. This PL/SQL procedure is called once for each row in the table to be indexed, and is passed the ROWID value of the current row being indexed. The actual contents of the procedure is entirely up to the owner, but it is normal to fetch data from one or more columns from database tables.
In both cases, we still need to take care of updates - making sure that we have all the triggers necessary to update the indexed column (and, in case 1, the summary table) whenever any of the data to be indexed gets changed.

I've written full examples of both these techniques, as SQL scripts to be run in the SQL*Plus tool. You will need to run them as a user who has CTXAPP role and CREATE DIRECTORY privilege.

Part of the data to be indexed is a Microsoft Word file called "1.doc". You should create this file in Word, preferably containing the single line of text: "test document". This file can be saved anywhere, but the SQL scripts need to be changed so that the "create or replace directory" command refers to the right location. In the example, I've used C:\doc.

multi_table_indexing_1.sql : creates a summary table containing all the data, and uses multi_column_datastore Download link / View in browser

multi_table_indexing_2.sql : creates "virtual" documents using a procedure as a user_datastore Download link / View in browser

Wednesday Jan 13, 2010

Indexing MediaWiki with SES

[Read More]



« February 2016