Minor Description
After having an Issue with compaction stopping for some time. It can be an issue that the LogCleaner will always clean into 1 Segment per sizegroup.
Usually the Log enforces a maximum distance between min and max offset in a LogSegment. If that Distance would be exeeded in maybeRoll() a new logsegment would be rolled. I assume this is because relative offset might be stored as integer. The LogCleaner OTOH is not going to roll a new LogSegment as its only ever using 1 Segment to clean into.
A lenghty discussion about this can be found in the slack community:
https://confluentcommunity.slack.com/archives/C49R61XMM/p1506914441000005
The observed stacktrace is as follows:
https://gist.github.com/brettrann/ce52343692696a45d5b9f4df723bcd14
I could imagin also enfocing Integer.MAX_VALUE as offset distance in
groupSegmentsBySize in the LogCleaner to make sure a Segment doesnt exeed this limit.