On 6/17/2011 10:54 AM, Daniel Shahaf wrote:
Thomas Harold wrote on Fri, Jun 17, 2011 at 10:31:43 -0400:
And if you have a choice of file systems for the repository to be stored
on, make sure that it's something which can deal with a few hundred
thousand tiny files. On Linux, I'd suggest going with ext4 over ext3.
While db/revs in a FSFS repository can have its revisions packed to
reduce the file count, the db/revprops folder still consists of 1 tiny
file for every revision in the project in a FSFS repository.
revprops/ is sharded.
And in 1.7 (including the recent 1.7.0-alpha1) it is packed, too.
Good. Another of the many reasons that we're looking forward to 1.7.
Even with the sharding, those little revprop files are causing us issues
during backups (hotcopy -> rdiff-backup). Being able to pack those
revprop files is going to make a big difference as the backup process
will only have to track 2000-2200 files instead of 30,000 to 50,000.
(We have a few long-lived repositories with up to 25k revisions. And I
just finished splitting a 22GB repository with 15-16k revs into a bunch
of smaller repositories. Now the nightly backup can look at doing a
hotcopy on only the repositories with changes in the last 5 days.)
