apache
ZOOKEEPER-2332: Fix server failed to start for empty txn log
When a new txn log is created, it encounters a failure before writing to the file header. The txn log file is empty at this moment, the server starts up again, reads the file, and fails to load the database because the header is not recognized.
The patch detects that the empty txn log is at the end of the database, and if the file is empty, it deletes the file and prompts the server to reboot to reload it.
When one of the txn logs in the middle of the database is empty, it will refuse to start up since it's a real corruption.
I am not sure it works on all filesystem if we are renaming and writing at same time.
Indeed. A key-value store may be needed to solve this issue better.
@kezhuw
Introduces lock file (say ".zookeeper.lock") to gain exclusive(cooperatively) access to zookeeper directory.
I like this approach to address ZOOKEEPER-1621. Would you mind creating a patch? I think it should not be covered with this one, rather we deal with it separately.
