[jira] [Created] (KAFKA-980) Crash during log recovery can cause full recovery to never run

["Blake Smith (JIRA)" <ji...@apache.org>]

Blake Smith created KAFKA-980:
---------------------------------

             Summary: Crash during log recovery can cause full recovery to never run
                 Key: KAFKA-980
                 URL: https://issues.apache.org/jira/browse/KAFKA-980
             Project: Kafka
          Issue Type: Bug
    Affects Versions: 0.7.1, 0.8
            Reporter: Blake Smith


After an unclean shutdown of the Kafka server, if the broker throws an unhandled exception during log recovery, the broker can get in a state where recovery never runs on a log file.

We saw this problem manifest in production and is summarized on the mailing list here: http://mail-archives.apache.org/mod_mbox/kafka-users/201307.mbox/%3CCAKSpikjgp2sW2ycuf86JrjtAPxWBp92OOEmigVed=u=JFoPvTA@mail.gmail.com%3E

Because recovery state is not tracked explicitly, our kafka broker started writing data even when the log files were not fully recovered. It feels to me like a separate state flag for recovery should also be tracked in cases where recovery does not fully run. What do you guys think?

Steps to reproduce:

1. Shutdown the kafka broker
2. Create a directory named 'bogus' under the kafka log directory (won't parse since it has no partition number)
3. Remove .kafka_cleanshutdown from the log directory to force a recovery
4. Start the kafka broker, observe:
    - Recovery will run on partition segments until it reaches the bogus directory
    - Exception will be thrown during log loading from the bogus directory
    - Kafka will initiate a clean shutdown after the exception is thrown
5. Once the Kafka server is cleanly shutdown, start it again, observe:
    - Recovery will not try to run, since kafka was shutdown cleanly
    - Some partition log files have never been recovered
6. Remove the bogus log directory
7. Start Kafka broker, observe:
    - Recovery will not run
    - Kafka will start cleanly and begin accepting writes again, even though recovery has never run and logs might be in a corrupt state


--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira