Cleanup of unreferenced KahaDB journal log files data-<id>.log will occur every 30seconds by deafault. If a data file is in-use it will not be cleaned up.
The trace level logging of the org.apache.activemq.store.kahadb.MessageDatabase class provides insight into the cleanup process and will allow you to determine why a given data file is considered in-use and as a result, not a candidate for cleanup.
To debug, add the following (or similar) to your log4j.properties file (if needed):
Either restart AMQ and let the cleanup process run (give it a minute or two for example) or alternatively apply this logging configuration to a running broker via JMX. The "Broker" MBean exposes an operation called "reloadLog4jProperties" in JMX that can be used to tell the broker to reload its log4j.properties. Often its enough to apply this logging configuration for 2-5 minutes and then analyze the broker's log file.
Examine the log file and look for cleanup of the data files. The process starts with the complete set of known data files and queries the index on a per destination basis to prune this list. Anything that remains is a candidate for cleanup.
At some point you'll hit a destination and the number of data file ids will suddenly drop because that destination references them. It could be a DLQ or an offline durable subscriber.
Here is a quick sample
We get one candidate, data-87.log from the existing set of journal data files [86, 87, 163, 164]. There is a current transaction using 164, destination (Queue named E) '0\:E' has some messages in 163, destination '0:I' has messages in 86 and 87 is unreferenced. In this case, there must be some long standing unacked messages or a very slow consumer on destination '0:I'.
The '0:' prefix is shorthand for a queue, '1:' for a topic, i.e: dest:1:B is a topic named B.