Øystein Grøvlen (JIRA) wrote:
[ http://issues.apache.org/jira/browse/DERBY-298?page=comments#action_66168
]
Øystein Grøvlen commented on DERBY-298:
---------------------------------------
Looking at the code, I became a bit confused about the definition of an empty
log file. Scan.getNextRecordForward contains debug output when it detects an
empty log file. It will then return without setting knownGoodLogEnd. Hence,
new log records will be written to the end of the previous file. As Suresh
says this is probably to be able to handle crashes during log switch.
However, this is not what happens when I run the recovery part of the example in this
report. Since, currentLogFileLength is a large number, it detects "zapped log end
on log file", goes on to the next file, which does not exist, and returns. (Who
sets the length of a log file? Is this maximum size until a log switch is performed?)
The effect is the same, but this can not be used to detect an empty log file and apply
the solution proposed by Suresh. Instead, one would have to do some hairy file handling
at a later stage.
derby has two types of log files , one that works in RWS mode with a
preallocated log file ,
and one which uses file sync with out preallocation. In case of
preallocation , zeros are written to the log file to the length
specified by the logSwitchInterval (Default is 1 MB) , it is also
configurable by the user.
Empty log file can not be identified based on the length alone. . Only
way to declare a log file is empty/fresh is when No log records found
in the file during recover scan. As u noticed because it is a
preallocated file , when scan finds zeros on first after zeros, it
declares that there are no more log records. Any fix for this problem
has to handle both preallocated and non preallocated case .
Actually, I don't like idea of creating a new log file on boot even
for special conditions,
because spending more time than required does not make out users happy
! If you can find a fix that does not require a new log creation, it
will be great.
An alternative way to fix this would be to just create a dummy log record in
the new log file as part of the backup command. This would make the redo scan
end in the new log file. However, this will not work for those who do backup
with OS-commands (i.e., copy the files directly).
Backup with OS command should not be a problem , because there is no
support to perform roll forward recovery with these type of backups. If
it is a just a plain backup , it does not matter how logs are
written after the backup, because they are never used to do a restore.
This problem can also be fixed by writing dummy log record.
Filelogger.redo() code has to be fixed to understand this, currently it
looks at the logEnd only when a good log records is read.
I would also think it should be possible to do the log switch in such a way
that it is possible to detect during recovery whether the log switch had
completed or not. If this was the case, one could just set knownGoodLogEnd of
the redo scan to the start of the empty file if the log switch was completed.
Does anyone know if this is possible?
Yes. It should be, with some changes to do redo/scan code.
Another possible solution I was thinking to identify whether a log
switch is good or not is by writing a INT (4 bytes of zeros ) after
log file initialization. As 512 bytes writes suppose to be atomic. if
(log file length > = LOG_FILE_HEADER_SIZE(24) +4) then the log switch
before the crash can be fixed as good one and fix the scan code to
use the empty log file instead of writing to the previous log file.
Thanks
-suresht