Bazaar Subversion Plugin

Merge lp://staging/~maxb/bzr-svn/cache-dev into lp://staging/bzr-svn/1.0

cache-dev
Merge into 1.0

Proposed by Max Bowsher on 2011-01-19

Status:	Superseded
Proposed branch:	lp://staging/~maxb/bzr-svn/cache-dev
Merge into:	lp://staging/bzr-svn/1.0
Diff against target:	159 lines (+21/-14) 5 files modified cache/__init__.py (+3/-0) cache/sqlitecache.py (+8/-2) cache/tdbcache.py (+0/-3) repository.py (+10/-6) revmeta.py (+0/-3)
To merge this branch:	bzr merge lp://staging/~maxb/bzr-svn/cache-dev
Related bugs:	Link a bug report

Reviewer	Review Type	Date Requested	Status
bzr-svn developers		2011-01-19	Pending
Review via email: mp+46761@code.staging.launchpad.net

This proposal has been superseded by a proposal from 2011-01-22.

Description of the change

Hi,

A trio of tweaks to the cache commit logic:

1. Since bzr-svn doesn't really care whether it's cache store is transactional or not, ideally this wouldn't need to be exposed in its interface more than is necessary. Therefore, move all of the period commit ("commit_conditionally") calls inside sqlitecache.py - which also makes sure that some code paths don't fail to call commit_conditionally when mutating the cache.

2. I found that bzr-svn was spending an awful lot of time doing "discovering revprop revisions" during many operations on some of my branches. It turns out that the "last checked revnum" watermark for the revid map cache was never being committed to the cache, so bzr-svn was needlessly reprocessing half a million revisions each time it was run. So, commit unconditionally when updating this watermark.

3. The previous thing got me thinking - if bzr-svn is throwing away work done because it's forgetting to commit, that's bad. So, looking around, it seems reasonable to commit during ForeignRepository.unlock().

Unmerged revisions

3511. By Max Bowsher on 2011-01-16: Add additional muttering in sqlite RevisionIdMapCache.
3510. By Max Bowsher on 2011-01-16: Remove vestigial method.
3509. By Max Bowsher on 2011-01-18: Commit the sqlite cache during ForeignRepository.unlock.
3508. By Max Bowsher on 2011-01-18: Updating the last_revnum_checked marks then end of a potentially arbitrarily
large amount of work - building the revid map for a project's entire history -
so commit it immediately. (I experienced this work going uncommitted and being
lost.)
3507. By Max Bowsher on 2011-01-18: Move commit_conditionally calls inside the cache implementation. This avoids
there being code-paths that never directly commit their caching work
(DiskCachingRevidMap and DiskCachingParentsProvider).
3506. By Jelmer Vernooij on 2011-01-16: mark as compatibler with bzr 2.4.
3505. By Jelmer Vernooij on 2011-01-15: Avoid using deprecated code.
3504. By Jelmer Vernooij on 2011-01-13: Remove accidentally committed pdb.
3503. By Jelmer Vernooij on 2011-01-12: Store short testament rather than long testament in bzr:testament
revision property.
3502. By Jelmer Vernooij on 2011-01-12: Update NEWS.