Skip to content

Releases: linkedin/kafka

Clearing the ZkReplicaStateMachine request batch state upon ControllerMovedException

09 Feb 23:44
0397da2
Compare
Choose a tag to compare
2.4.1.12

[LI-HOTFIX] Clearing the ZkReplicaStateMachine request batch state up…

Add broker info log when getting UNKNOWN_TOPIC_OR_PARTITION error

09 Feb 23:50
164b7d6
Compare
Choose a tag to compare
2.3.0.22

[LI-HOTFIX] Add broker info log when getting UNKNOWN_TOPIC_OR_PARTITI…

Controlled shutdown safety check

03 Feb 00:39
32a2f98
Compare
Choose a tag to compare

Adds a topic availability safety check to the controlled shutdown protocol.

New broker config properties include:

  • controlled.shutdown.safety.check.enable: Enable this safety check for any ControlledShutdownRequest handled by the controller.
  • controlled.shutdown.safety.check.redundancy.factor: Additional replicas above min.insync.replicas that are required for any topic-partition to allow shutdown of a broker. Default is 2, so if a topic has min.insync.replicas=2, then the ISR must have 4 replicas to allow shutdown.

Also included in this release is a fix from upstream (10b0757) for KAFKA-10794, Replica leader election is too slow in the case of too many partitions.

Move Logging to companion classes

07 Jan 18:51
dd0227e
Compare
Choose a tag to compare
[LI-HOTFIX] Move the logging trait to a companion object. (#104)

TICKET = KAFKA-10877
LI_DESCRIPTION = When Log4j2 is used in JDK11 it uses a different class from JDK 8 to determine where the logger was instantiated (which class instantiated the logger). The StackWalker class is used to do this. The StackWaker turns out to be very slow which causes low RHIPR on the brokers. This cpu cost occurs even if a log message is not emitted as the logger needs to perform this initialization even to find out if a log message needs to be emitted. If loggers are not initialized frequently then this cost is low when amortized over runtime of the broker. This patch moves the Logging trait from some classes into companion objects. This initializes the logger only once for each of these classes instead of every time a class is instantiated. The larger patch of doing this move to every class which currently mixes in Logger is more complicated and will be attempted in open source.

EXIT_CRITERIA = TICKET [KAFKA-10877]

Test uploading archives and ignored flaky test

06 Jan 18:42
4a53cf7
Compare
Choose a tag to compare

Test uploading archives and ignored flaky test

Adding BINTRAY_USER and BINTRAY_KEY for uploading archives

06 Jan 02:15
d16d4d2
Compare
Choose a tag to compare

Adding BINTRAY_USER and BINTRAY_KEY for uploading archives

Testing the Github Actions uploadArchives workflow

05 Jan 18:25
4ac92ac
Compare
Choose a tag to compare

Testing the Github Actions uploadArchives workflow

Switching the CI from Travis to Github Actions

05 Jan 17:58
bb6cab9
Compare
Choose a tag to compare

Switching the CI from Travis to Github Actions

Prevent UNKNOWN_TOPIC_OR_PARTITION for all topics metadata request

02 Dec 21:57
04b5686
Compare
Choose a tag to compare
  • Prevent UNKNOWN_TOPIC_OR_PARTITION for all topics metadata request
  • Generate logs to help estimate the amount of data loss during ULE

Reject stale metadata from a different cluster

29 Oct 21:43
c7b3619
Compare
Choose a tag to compare

Include fix to reject stale metadata from a different cluster.