Master-slave vs. peer-to-peer archictecture: benefits and problems
By Ashok_Ora on Jun 09, 2014
Almost two decades ago, I was a member of a database development team that introduced adaptive locking. Locking, the most popular concurrency control technique in database systems, is pessimistic. Locking ensures that two or more conflicting operations on the same data item don’t “trample” on each other’s toes, resulting in data corruption. In a nutshell, here’s the issue we were trying to address. In everyday life, traffic lights serve the same purpose. They ensure that traffic flows smoothly and when everyone follows the rules, there are no accidents at intersections.
As I mentioned earlier, the problem with typical locking protocols is that they are pessimistic. Regardless of whether there is another conflicting operation in the system or not, you have to hold a lock! Acquiring and releasing locks can be quite expensive, depending on how many objects the transaction touches. Every transaction has to pay this penalty. To use the earlier traffic light analogy, if you have ever waited at a red light in the middle of nowhere with no one on the road, wondering why you need to wait when there’s clearly no danger of a collision, you know what I mean.
The adaptive locking scheme that we invented was able to minimize the number of locks that a transaction held, by detecting whether there were one or more transactions that needed conflicting eyou could get by without holding any lock at all. In many “well-behaved” workloads, there are few conflicts, so this optimization is a huge win. If, on the other hand, there are many concurrent, conflicting requests, the algorithm gracefully degrades to the “normal” behavior with minimal cost.
We were able to reduce the number of lock requests per TPC-B transaction from 178 requests down to 2! Wow! This is a dramatic improvement in concurrency as well as transaction latency.
The lesson from this exercise was that if you can identify the common scenario and optimize for that case so that only the uncommon scenarios are more expensive, you can make dramatic improvements in performance without sacrificing correctness.
So how does this relate to the architecture and design of some of the modern NoSQL systems? NoSQL systems can be broadly classified as master-slave sharded, or peer-to-peer sharded systems. NoSQL systems with a peer-to-peer architecture have an interesting way of handling changes. Whenever an item is changed, the client (or an intermediary) propagates the changes synchronously or asynchronously to multiple copies (for availability) of the data. Since the change can be propagated asynchronously, during some interval in time, it will be the case that some copies have received the update, and others haven’t.
What happens if someone tries to read the item during this interval? The client in a peer-to-peer system will fetch the same item from multiple copies and compare them to each other. If they’re all the same, then every copy that was queried has the same (and up-to-date) value of the data item, so all’s good. If not, then the system provides a mechanism to reconcile the discrepancy and to update stale copies.
So what’s the problem with this? There are two major issues:
First, IT’S HORRIBLY PESSIMISTIC because, in the common case, it is unlikely that the same data item will be updated and read from different locations at around the same time! For every read operation, you have to read from multiple copies. That’s a pretty expensive, especially if the data are stored in multiple geographically separate locations and network latencies are high.
Second, if the copies are not all the same, the application has to reconcile the differences and propagate the correct value to the out-dated copies. This means that the application program has to handle discrepancies in the different versions of the data item and resolve the issue (which can further add to cost and operation latency).
Resolving discrepancies is only one part of the problem. What if the same data item was updated independently on two different nodes (copies)? In that case, due to the asynchronous nature of change propagation, you might land up with different versions of the data item in different copies. In this case, the application program also has to resolve conflicts and then propagate the correct value to the copies that are out-dated or have incorrect versions. This can get really complicated. My hunch is that there are many peer-to-peer-based applications that don’t handle this correctly, and worse, don’t even know it. Imagine have 100s of millions of records in your database – how can you tell whether a particular data item is incorrect or out of date? And what price are you willing to pay for ensuring that the data can be trusted? Multiple network messages per read request? Discrepancy and conflict resolution logic in the application, and potentially, additional messages? All this overhead, when all you were trying to do was to read a data item.
Wouldn’t it be simpler to avoid this problem in the first place? Master-slave architectures like the Oracle NoSQL Database handles this very elegantly. A change to a data item is always sent to the master copy. Consequently, the master copy always has the most current and authoritative version of the data item. The master is also responsible for propagating the change to the other copies (for availability and read scalability). Client drivers are aware of master copies and replicas, and client drivers are also aware of the “currency” of a replica. In other words, each NoSQL Database client knows how stale a replica is.
This vastly simplifies the job of the application developer. If the application needs the most current version of the data item, the client driver will automatically route the request to the master copy. If the application is willing to tolerate some staleness of data (e.g. a version that is no more than 1 second out of date), the client can easily determine which replica (or set of replicas) can satisfy the request, and route the request to the most efficient copy. This results in a dramatic simplification in application logic and also minimizes network requests (the driver will only send the request to exactl the right replica, not many).
So, back to my original point. A well designed and well architected system minimizes or eliminates unnecessary overhead and avoids pessimistic algorithms wherever possible in order to deliver a highly efficient and high performance system. If you’ve every programmed an Oracle NoSQL Database application, you’ll know the difference!