Jeff Nadler created KAFKA-9899: ---------------------------------- Summary: LogCleaner Tries To Clean Single Partition Over 1000x/Minute Key: KAFKA-9899 URL: https://issues.apache.org/jira/browse/KAFKA-9899 Project: Kafka Issue Type: Bug Components: log cleaner Affects Versions: 2.4.1 Environment: ubuntu bionic, openjdk11.0.6, kafka 2.4.1 Reporter: Jeff Nadler Attachments: CPU-Usage.png
I had previously believed this to be the same issue as KAFKA-8764, but I took a closer look when it persisted after upgrading to 2.4.1 and now believe this is a different bug. For a topic that is a very low traffic, compact topic the log cleaner will sometimes - for a period of usually 2 hours or longer - get stuck in a loop where it tries to clean the same partition for the same offset range nonstop, and the log cleaner thread consumes 100% of a single core during this time. h4. 1396 attempts in a single minute: {{root@stage-obs-kafka01:/var/log/kafka# cat log-cleaner.log | grep 22:22: | grep "offset range" | wc -l}} {{1396}} h4. All 1396 of these are looking at the same partition and same offset range: {{[2020-04-21 22:22:59,862] INFO Cleaner 0: Building offset map for log elauneind-firebolt-messages-sfo-0 for 0 segments in offset range [22943108, 22912825). (kafka.log.LogCleaner)}} These attempts are separated by on average only 30ms. This is a small 3 node cluster, note that the CPU graph attached is very clearly bimodal for each node: low when the log cleaner is not "stuck", and much higher when it is. Eventually the log cleaner appears to find a segment to clean (because enough traffic has arrived?) and the loop is broken... for a time. Note that it finds "1 segments" and then finally moves on to check other topic-partitions. {{...tens of thousands of this first one then}} {{[2020-04-21 20:06:02,531] INFO Cleaner 0: Building offset map for log elauneind-firebolt-messages-sfo-0 for 0 segments in *offset range* [23591841, 23575583). (kafka.log.LogCleaner)}}{{[2020-04-21 20:06:02,567] INFO Cleaner 0: Building offset map for log elauneind-firebolt-messages-sfo-0 for 1 segments in *offset range* [23591841, 23621641). (kafka.log.LogCleaner)}}{{[2020-04-21 20:43:04,309] INFO Cleaner 0: Building offset map for log elauneind-firebolt-messages-s2r1-0 for 1 segments in *offset range* [2687968, 2732498). (kafka.log.LogCleaner)}} h4. The topic gets about 100 messsages/minute, and it's config is: {{Topic: elauneind-firebolt-messages-sfo PartitionCount: 1 ReplicationFactor: 3 Configs: min.insync.replicas=1,cleanup.policy=compact,delete,segment.bytes=102400000,retention.ms=900000,message.format.version=2.3-IV1,min.compaction.lag.ms=300000,min.cleanable.dirty.ratio=0.2,unclean.leader.election.enable=true,retention.bytes=1073741824}}{{ Topic: elauneind-firebolt-messages-sfo Partition: 0 Leader: 0 Replicas: 0,2,1 Isr: 0,1,2}} -- This message was sent by Atlassian Jira (v8.3.4#803005)