Details
Major Description
So this started with me thinking this was a mirrormaker2 issue because here are the symptoms I am seeing:
I'm encountering an odd issue with mirrormaker2 with our remote replication setup to high latency remote sites (satellite).
Every few days we get several topics completely re-replicated, this appears to happen after a network connectivity outage. It doesn't matter if it's a long outage (hours) or a short one (minutes). And it only seems to affect a few topics.
I was finally able to track down some logs showing the issue. This was after an hour-ish long outage where connectivity went down. There were lots of logs about connection timeouts, etc. Here is the relevant part when the connection came back up:
2023-09-08 16:52:45,380 INFO [scbi->gcp.MirrorSourceConnector|worker] [AdminClient clientId=mm2-admin-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector|replication-source-admin] Disconnecting from node 0 due to socket connection setup timeout. The timeout value is 63245 ms. (org.apache.kafka.clients.NetworkClient) [kafka-admin-client-thread | mm2-admin-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector|replication-source-admin] 2023-09-08 16:52:45,380 INFO [scbi->gcp.MirrorSourceConnector|worker] [AdminClient clientId=mm2-admin-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector|replication-source-admin] Metadata update failed (org.apache.kafka.clients.admin.internals.AdminMetadataManager) [kafka-admin-client-thread | mm2-admin-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector|replication-source-admin] 2023-09-08 16:52:47,029 INFO [scbi->gcp.MirrorSourceConnector|task-1] [Consumer clientId=mm2-consumer-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector-1|replication-consumer, groupId=null] Disconnecting from node 0 due to socket connection setup timeout. The timeout value is 52624 ms. (org.apache.kafka.clients.NetworkClient) [task-thread-scbi->gcp.MirrorSourceConnector-1] 2023-09-08 16:52:47,029 INFO [scbi->gcp.MirrorSourceConnector|task-1] [Consumer clientId=mm2-consumer-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector-1|replication-consumer, groupId=null] Error sending fetch request (sessionId=460667411, epoch=INITIAL) to node 0: (org.apache.kafka.clients.FetchSessionHandler) [task-thread-scbi->gcp.MirrorSourceConnector-1] 2023-09-08 16:52:47,336 INFO [scbi->gcp.MirrorSourceConnector|worker] refreshing topics took 67359 ms (org.apache.kafka.connect.mirror.Scheduler) [Scheduler for MirrorSourceConnector: scbi->gcp|scbi->gcp.MirrorSourceConnector-refreshing topics] 2023-09-08 16:52:48,413 INFO [scbi->gcp.MirrorSourceConnector|task-1] [Consumer clientId=mm2-consumer-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector-1|replication-consumer, groupId=null] Fetch position FetchPosition{offset=4918131, offsetEpoch=Optional[0], currentLeader=LeaderAndEpoch{leader=Optional[kafka.scbi.eng.neoninternal.org:9094 (id: 0 rack: null)], epoch=0}} is out of range for partition reading.sensor.hfp01sc-0, resetting offset (org.apache.kafka.clients.consumer.internals.AbstractFetch) [task-thread-scbi->gcp.MirrorSourceConnector-1] (Repeats for 11 more topics) 2023-09-08 16:52:48,479 INFO [scbi->gcp.MirrorSourceConnector|task-1] [Consumer clientId=mm2-consumer-scbi|scbi->gcp|scbi->gcp.MirrorSourceConnector-1|replication-consumer, groupId=null] Resetting offset for partition reading.sensor.hfp01sc-0 to position FetchPosition{offset=3444977, offsetEpoch=Optional.empty, currentLeader=LeaderAndEpoch{leader=Optional[kafka.scbi.eng.neoninternal.org:9094 (id: 0 rack: null)], epoch=0}}. (org.apache.kafka.clients.consumer.internals.SubscriptionState) [task-thread-scbi->gcp.MirrorSourceConnector-1] (Repeats for 11 more topics)
The consumer reports that offset 4918131 is out of range for this topic/partition, but that offset still exists on the remote cluster. I can go pull it up with a consumer right now. The earliest offset in that topic that still exists is 3444977 as of yesterday. We have 30 day retention configured so pulling in 30 days of duplicate data is very not good. It almost seems like a race condition as there are 38 topics we replicate but this only affected 12 (on this occurance).
The number of topics affected seems to vary each time. Today I see one site has 2 topics it is resending, and another has 13.
But since opening this issue originally what I have discovered is that this error occurs every time I have a power failure at a remote site I mirror to. I can reproduce this issue by doing a hard reset on my single node broker setup. The broker comes back up cleanly after processing unflushed messages to segments, but while it's coming up I consistently have an issue where I get these partition out of range errors from mirrormaker2. I've looked around online and found a few other issues that sound similar to mine involving kafka cluster upgrades, and cpu hangs, which sounds like a similar state (crashing broker, unclean shutdown). Something about unclean recovery is causing this to occur.