How Do I Prevent Blocking Locks in Greenplum?

This is the fifth in a series of “How do I do X in Greenplum database” and today’s topic is How do I prevent “blocking locks” in Greenplum?

Since different databases handle read consistency and locking differently, I will discuss how Oracle, Microsoft SQL Server, PostgreSQL, and finally Greenplum handle locking. This will give you a good way to compare and contrast the different products.

Oracle handles read consistency with using “rollback”. When you update a table, the row is overwritten but first, the old version of the data is put into a rollback segment and the row is marked with a timestamp and a pointer to the old version. This allows users to query the database and not be blocked while an update happens.

If a query started at time1 and and update happens at time2, the query will see a snapshot of the data as of time1. To make this work, the query follows the pointer to the rollback segment to get the old version of the rows that were updated.

Oracle Problems
There are a couple of problems with this solution. The first is Disk IO because the database has to first read the data file and then read the rollback segment.

The second problem with this is the rollback segment must be sized pretty large in a warehouse environment when you can easily have a query execute for a long period of time. You will eventually run out of rollback and get the “snapshot too old” error on long running queries if you are actively changing data.

SQL Server
SQL Server uses locking by default to provide read consistency. This means an update or even an insert can block other sessions from reading a table. Yuck. Starting with SQL Server 2005, you can enable Read Committed Snapshot Isolation (RCSI) which makes the database behave more like Oracle but it uses TempDB instead of rollback for the old version of the rows updated.

SQL Server Problems
To get around locking, developers will use “read uncommitted” which is a dirty read. I never use a dirty read because it is a hack. Don’t do it.

Using RCSI uses TempDB so more IO is involved. Unfortunately, I also almost never see it used in companies. Most SQL Server shops don’t even know this option is available so they continue with their dirty reads.

PostgreSQL uses Multi-Version Concurrency Control (MVCC) to handle read consistency. In a nutshell, readers don’t block writers and writers don’t block readers. Each transaction sees a snapshot of the database rather than locking tables.

To a user, PostgreSQL behaves similarly to Oracle but it doesn’t use rollback. Instead, an update adds a new row version to the existing table. Because PostgreSQL doesn’t use rollback, you will never get a “snapshot too old” error on long running queries.

PostgreSQL Problems
The downside to PostgreSQL happens when an update affects a large percentage of a table. If you update an entire table, it effectively doubles the size on disk. To make the old rows available again, a “vacuum” will need to be run so that new updates and/or inserts will use the deleted rows. PostgreSQL uses a background daemon to automatically vacuum tables so this usually isn’t a big deal.

Greenplum is a fork of PostgreSQL 8.2 so it uses MVCC to handle read consistency. Readers don’t block writers and writers don’t block readers. Each transaction sees a snapshot of the database rather than locking tables.

Greenplum MVCC Exception
Greenplum differs from PostgreSQL for update and delete commands. It acquires an Exclusive Lock on a table. However, this lock doesn’t block readers. SQL queries will return data and users will see a snapshot of the data that isn’t dirty. This Exclusive Lock does however block other update and delete commands which is different from PostgreSQL.

Greenplum Database provides read consistency with Multi-Version Concurrency Control (MVCC) similarly to PostgreSQL but update and delete operations are serialized. You don’t need to worry about blocking locks for selecting data because MVCC uses snapshots of the database instead of blocking locks. You can get a blocking lock on update and delete when another update or delete command is executing on the same table but readers aren’t blocked! You will likely never even notice this lock because the database will handle the updates and deletes one at a time while you can still select data.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.