On Sun, Apr 6, 2014 at 5:36 PM, Ivan Zhakov <i...@visualsvn.com> wrote:
> On 6 April 2014 18:31, Stefan Fuhrmann <stefan.fuhrm...@wandisco.com> > wrote: > > On Wed, Apr 2, 2014 at 12:13 PM, Ivan Zhakov <i...@visualsvn.com> wrote: > >> > >> On 31 January 2014 14:57, Ivan Zhakov <i...@visualsvn.com> wrote: > >> > On 31 January 2014 05:50, Evgeny Kotkov <evgeny.kot...@visualsvn.com> > >> > wrote: > >> >>> This only affects non-sharded repositories with rev-local IDs, > >> >>> i.e. those in SVN 1.4 format. For those, it is writing 3 files > >> >>> instead of 2 per rev now. > >> >> > >> >> I assume you are talking about r1560723 [1]. If I read the code > >> >> correctly... > >> >> [[[ > >> >> if ( (!max_files_per_dir || rev % max_files_per_dir == 0) > >> >> && dst_ffd->format >= SVN_FS_FS__MIN_NO_GLOBAL_IDS_FORMAT) > >> >> SVN_ERR(hotcopy_update_current(&dst_youngest, dst_fs, rev, > >> >> iterpool)); > >> >> ]]] > >> >> > >> >> ...hotcopy now checkpoints after every revision for all non-sharded > >> >> repositories > >> >> with FSFS format >= 3. So, checkpointing happens for all FSFS format > >> >> 1/2 > >> >> repositories upgraded via 'svnadmin upgrade' (with linear layout), > that > >> >> were > >> >> not fsfs-reshard'ed or dump/loaded into a repository with newer > format. > >> >> > >> >> I am not aware of how many people reshard or dump/load their old > >> >> repositories > >> >> after upgrade, but I did a quick benchmark for a real-world > repository > >> >> and on my > >> >> machine it shows 7x performance degradation with checkpointing > enabled. > >> >> Is it > >> >> worth the ability to re-run the backup from a checkpoint upon > >> >> cancellation? > >> >> > >> >> [[[ > >> >> # svnrdump http://googletest.googlecode.com/svn/trunk/ > >> >> # load the dump into a --compatible-version=1.3 repository > >> >> # upgrade the repository with the most recent svnadmin > >> >> > >> >> # disable checkpointing, benchmark making 100 hotcopy backups: > >> >> real 0m18.741s > >> >> user 0m2.552s > >> >> sys 0m13.432s > >> >> > >> >> # now enable checkpointing and repeat the benchmark: > >> >> real 2m5.793s > >> >> user 0m0.836s > >> >> sys 0m34.840s > >> >> ]]] > >> >> > >> >> [1] http://svn.apache.org/viewvc?view=revision&revision=r1560723 > >> >> > >> > That's what I suspected. In this case I think r1560723 should not be > >> > backported to 1.8.x for the following reasons: > >> > 1. Significant performance degradation for 'svnadmin hotcopy' with > >> > '--incremental' flag. > >> > 2. The change doesn't fix original circular dependency problem in FSFS > >> > 3. 'svnadmin recover' writes fake value to CURRENT file which very bad > >> > practice IMHO and could lead undefined consequences. > >> > > >> Hi Stefan, > >> > >> Is this problem fixed or it's better to revert r1560723 for reasons I > >> stated above? > > > > > > Hi Ivan, > > > > Evgeny applied his patch a month ago as r1573744. > > > Hi Stefan, > > Yes, I know. But svnadmin hotcopy performance regression was not fixed > as far I know. > Oh - sorry! When I read your post, I didn't expand the quote. I don't have string feelings either way. As it makes laggards happy, I'm fine with the revert. Reverted in r1585516. -- Stefan^2.
