Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
5.3.0
-
None
-
RHEL 5, Linux 2.6.18, PostgreSQL 8.3.4, java 1.6.0_05-b13
-
Patch Available
Description
Hi all,
I made the following improvements to the JDBC persistence adapter:
1. Implemented PostgresqlJDBCAdapter.doDeleteOldMessages() method that is at least three orders of magnitude faster than the default implementation. The same optimization may apply to other databases so you may want to consider moving it to DefaultJDBCAdapter.
2. Changed DefaultJDBCAdapter.doRecoverNextMessages() to always process at least 1000 messages at a time. This results in a huge performance improvement. However, this implementation is a hack (see comment). Somebody may want to look into why maxReturned gets set to 1 on the third call to doRecoverNextMessages(). I was able to consistently reproduce this behaviour.
3. Fixed JDBCPersistenceAdapter so that it does not double-call cleanup() [once in the main thread, once in the worker thread], and also so that it does not hang ActiveMQ initialization when there are lots of pending messages.
Performance is now limited by repeated calls to JDBCTopicMessageStore.acknowledge(). This causes an update for every single message (updates last_acked_id in activemq_acks table). I don't know enough about ActiveMQ architecture to optimize this, but perhaps someone should look into this. For example, if it only did this update after processing a batch of messages (say 1000 at a time), it would be a lot faster. The current implementation keeps incrementing last_acked_id, once per delivered message.
Diffs attached.
Eugene
Here is original post on the developer forum: http://www.nabble.com/JDBC-persistence-adapter-improvements-td21086330.html#a21086330