Doesn't the wide-area HBase replication use the .oldlogs to keep the slave
hBase cluster in sync?
thanks
dhruba
On Tue, Jul 5, 2011 at 7:27 AM, Lars George <lars.geo...@gmail.com> wrote:
> Hi,
>
> Ah, I see Ted has that also questioned in HBASE-4010. Good.
>
> And I was slightly wrong below, as there is a .logs and .oldlogs for the
> newly created regions. I agree that when there is no chance for the region
> to take edits before it gets the proper log, then we should drop this. Those
> dangling dirs do not harm, but also do not help either.
>
> Lars
>
> On Jul 5, 2011, at 8:39 AM, Lars George wrote:
>
> > Hi,
> >
> > Could someone confirm that we dropped the replay of the .oldlogs in the
> region directory because it is not needed anymore? Before the master rewrite
> we were, just as now, creating the region and for a brief period it had its
> own log directory, the .oldlogs under the region directory. The
> HRegion.initialize() would replay it in case an edit to the region arrived
> in between. It would then also remove it:
> >
> > public void initialize(Path initialFiles, final Progressable reporter)
> > throws IOException {
> > Path oldLogFile = new Path(regiondir, HREGION_OLDLOGFILE_NAME);
> > ...
> > // Play log if one. Delete when done.
> > doReconstructionLog(oldLogFile, minSeqIdToRecover, maxSeqId,
> reporter);
> > if (fs.exists(oldLogFile)) {
> > if (LOG.isDebugEnabled()) {
> > LOG.debug("Deleting old log file: " + oldLogFile);
> > }
> > fs.delete(oldLogFile, false);
> > }
> > ...
> >
> > Now we still create that directory, but the above code has been dropped.
> Only the recovered.edits is replayed instead while the .oldlogs now hang
> around. I am not even sure when they may contain an entry, but if they don't
> then we can either do not use "<regiondir>/.oldlogs", but point it to the
> new "<regiondir>/.tmp" - or we can add a unconditional delete in the
> initialize as before.
> >
> > Create a JIRA for this?
> >
> > Lars
>
>
--
Connect to me at http://www.facebook.com/dhruba
