[
https://issues.apache.org/jira/browse/KAFKA-2303?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Guozhang Wang updated KAFKA-2303:
---------------------------------
Fix Version/s: 0.8.3
> Fix for KAFKA-2235 LogCleaner offset map overflow causes another compaction
> failures
> ------------------------------------------------------------------------------------
>
> Key: KAFKA-2303
> URL: https://issues.apache.org/jira/browse/KAFKA-2303
> Project: Kafka
> Issue Type: Bug
> Components: core, log
> Affects Versions: 0.8.2.1
> Reporter: Alexander Demidko
> Assignee: Jay Kreps
> Fix For: 0.8.3
>
>
> We have rolled out the patch for KAFKA-2235 to our kafka cluster, and
> recently instead of
> {code}
> "kafka.log.LogCleaner - [kafka-log-cleaner-thread-0], Error due to
> java.lang.IllegalArgumentException: requirement failed: Attempt to add a new
> entry to a full offset map."
> {code}
> we started to see
> {code}
> kafka.log.LogCleaner - [kafka-log-cleaner-thread-0], Error due to
> java.lang.IllegalArgumentException: requirement failed: 131390902 messages in
> segment <topic-name>-cgstate-8/00000000000079840768.log but offset map can
> fit only 80530612. You can increase log.cleaner.dedupe.buffer.size or
> decrease log.cleaner.threads
> {code}
> So, we had to roll it back to avoid disk depletion although I'm not sure if
> it needs to be rolled back in trunk. This patch applies more strict checks
> than were in place before: even if there is only one unique key for a
> segment, cleanup will fail if this segment is too big.
> Does it make sense to eliminate a limit for the offset map slots count, for
> example to use an offset map backed by a memory mapped file?
> The limit of 80530612 slots comes from memory / bytesPerEntry, where memory
> is Int.MaxValue (we use only one cleaner thread) and bytesPerEntry is 8 +
> digest hash size. Might be wrong, but it seems if the overall number of
> unique keys per partition is more than 80M slots in an OffsetMap, compaction
> will always fail and cleaner thread will die.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
