[svn.haxx.se] · SVN Dev · SVN Users · SVN Org · TSVN Dev · TSVN Users · Subclipse Dev · Subclipse Users · this month's index

Re: fsfs revprop packing in f5 Re: Does fsfs revprop packing no longer allow usage of traditional backup software?

From: Hyrum K Wright <hyrum_at_hyrumwright.org>
Date: Thu, 7 Jul 2011 08:48:38 -0500

On Wed, Jul 6, 2011 at 2:30 PM, Hyrum K Wright <hyrum_at_hyrumwright.org> wrote:
> On Wed, Jul 6, 2011 at 11:47 AM, Daniel Shahaf <danielsh_at_elego.de> wrote:
>> This thread is now the only non-FSFS release blocker (filed as #3944).
>> Last I checked there were at least three solutions suggested, but no
>> consensus on which solution to implement.
>> Some suggestions were
>> 0. Leave things as they are
>> 1. Allow packing revisions without packing revprops.
>>   (revprops/ remains as in 1.6/f4)
>> 2. Have all revprops in the DB all the time, never in plain files.
>> 3. Swap the DB for some other "A thousand revision's revprops in one
>>   file" solution. [several suggestions as to that file's format]
>> Can we decide on what to do here?
>> Thanks,
>> Daniel
>> ---------
>> My opinion:
>> * (1) is orthogonal to the others, but may be a good idea if we refactor
>>  the FS so shortly before the release
>> * (2) simplifies things, doesn't solve the problems with having an
>>  SQLite db authoritative for parts of the FS storage
>>  (read: cp(1) unsafe)
>> * (3) has my +1, assuming it's sufficiently performant and the concrete
>>  design is reasonable
>> * (0) would mean that if we refactor revprop storage later, 1.8 servers
>>  will have to try and read revprops from *three* places; and lots of
>>  headache in the upgrade (and read-from-a-being-upgraded-FS) codepaths.
>>  So, if f5 should be improved, I'd rather do that /before/ it's
>>  released (and has to be indefinitely supported).
> After a bit of thinking and discussion, Daniel and I have come up with
> what we think is an acceptable solution, and I'm posting it here for
> validation.  (Daniel, please correct me if I've gotten something
> wrong.)
> Revision properties will *not* be packed in an sqlite database, but
> will instead be packed in a single packfile, much like revision are to
> today.  The key difference is that instead of having a separate
> manifest file, the manifest will be prepended to the packfile, meaning
> the two can be atomically replaced in the case of a propedit.
> This solution has at least of couple of advantages:
>  * No need to check a separate "edited" file before reading the packfile
>  * The repo maintains consistency in the case of a filesystem copy
> (helpful for backups)
> Revprops wouldn't be packed until explicitly asked to do so by
> 'svnadmin pack' which means the frequent post-commit revprop editing
> wouldn't pose a performance problem.  In addition, the revprop
> packfile manifest information won't be cached, since the manifest may
> change.  We don't anticipate this to be a problem, since it only adds
> an extra seek() to the revprop lookup process (rather than the open()
> + seek() in the rev packing world).

I've created the revprop-packing branch to do the implementation. It
currently has a skeleton BRANCH-README, which Daniel will be
augmenting shortly.

Received on 2011-07-07 15:49:13 CEST

This is an archived mail posted to the Subversion Dev mailing list.