[KAFKA-3039] Temporary loss of leader resulted in log being completely truncated

• Type: Bug
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Affects Version/s: 0.9.0.0
• Component/s: core
• Labels: reliability
• Environment: Debian 3.2.54-2 x86_64 GNU/Linux
• Assignee: Unassigned
• Reporter: Imran Patel
• Created: 23/Dec/15 21:28
• Updated: 20/Jul/20 08:58

Description

We had an event recently where the temporarily loss of a leader for a partition (during a manual restart), resulted in the leader coming back with no high watermark state and truncating its log to zero. Logs (attached below) indicate that it did have the data but not the commit state. How is this possible?

Leader (broker 3)
[2015-12-18 21:19:44,666] INFO Completed load of log messages-14 with log end offset 14175963374 (kafka.log.Log)
[2015-12-18 21:19:45,170] INFO Partition [messages,14] on broker 3: No checkpointed highwatermark is found for partition [messages,14] (kafka.cluster.Partition)
[2015-12-18 21:19:45,238] INFO Truncating log messages-14 to offset 0. (kafka.log.Log)
[2015-12-18 21:20:34,066] INFO Partition [messages,14] on broker 3: Expanding ISR for partition [messages,14] from 3 to 3,10 (kafka.cluster.Partition)

Replica (broker 10)
[2015-12-18 21:19:19,525] INFO Partition [messages,14] on broker 10: Shrinking ISR for partition [messages,14] from 3,10,4 to 10,4 (kafka.cluster.Partition)
[2015-12-18 21:20:34,049] ERROR [ReplicaFetcherThread-0-3], Current offset 14175984203 for partition [messages,14] out of range; reset offset to 35977 (kafka.server.ReplicaFetcherThread)
[2015-12-18 21:20:34,033] WARN [ReplicaFetcherThread-0-3], Replica 10 for partition [messages,14] reset its fetch offset from 14175984203 to current leader 3's latest offset 35977 (kafka.server.ReplicaFetcherThread)

Some relevant config parameters:
offsets.topic.replication.factor = 3
offsets.commit.required.acks = -1
replica.high.watermark.checkpoint.interval.ms = 5000
unclean.leader.election.enable = false