Vamsi Subhash Achanta created KAFKA-3359:
--------------------------------------------
Summary: Parallel log-recovery of un-flushed segments on startup
Key: KAFKA-3359
URL: https://issues.apache.org/jira/browse/KAFKA-3359
Project: Kafka
Issue Type: Bug
Components: log
Affects Versions: 0.9.0.1, 0.8.2.2
Reporter: Vamsi Subhash Achanta
Assignee: Jay Kreps
Priority: Minor
On startup, currently the log segments within a logDir are loaded sequentially
when there is a un-clean shutdown. This will take a lot of time for the
segments to be loaded as the logSegment.recover(..) is called for every segment
and for brokers which have many partitions, the time taken will be very high
(we have noticed ~40mins for 2k partitions).
https://github.com/apache/kafka/pull/1035
This pull request will make the log-segment load parallel with two configurable
properties "log.recovery.threads" and "log.recovery.max.interval.ms".
Logic:
1. Have a threadpool defined of fixed length (log.recovery.threads)
2. Submit the logSegment recovery as a job to the threadpool and add the future
returned to a job list
3. Wait till all the jobs are done within req. time
(log.recovery.max.interval.ms - default set to Long.Max).
4. If they are done and the futures are all null (meaning that the jobs are
successfully completed), it is considered done.
5. If any of the recovery jobs failed, then it is logged and
LogRecoveryFailedException is thrown
6. If the timeout is reached, LogRecoveryFailedException is thrown.
The logic is backward compatible with the current sequential implementation as
the default thread count is set to 1.
PS: I am new to Scala and the code might look Java-ish but I will be happy to
modify the code review changes.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
