On 7/27/07, Blair Zajac <blair@orcaware.com> wrote:
> Ivan Zhakov wrote:
> > On 7/26/07, Lieven Govaerts <svnlgo@mobsol.be> wrote:
> >> Well, the buildbot is pretty much useless when one buildslave is failing
> >> all the time, so why don't we revert it for now so you can take your
> >> time in fixing it?
> > Problem is what we have to revert to fix builds. Because it started
> > failing after I removed APR iconv dependency in commit r25650, but
> > actual problem is in commit r25430. And my commit only shows problem
> > in this commit, because made Subversion running faster :)
> > Personally I don't like idea to revert my commit (I don't have idea
> > about Blair's commit r25430).
> > I think that possible way is temporary fix commit r25430 with
> > including revision to transaction name.
>
> I was under the impression that the failure in the tests discussed here was the
> previously discussed issue with the lack of granularity of the clock under
> Windows causing the txn_names_are_not_reused() test to fail, but after looking
> at the logs, the two failures are not directly related to the clock:
>
> FAIL: fs-test.exe 22: check old revisions
> FAIL: fs-test.exe 28: test svn_fs_check_related
>
> The new txn txn_names_are_not_reused() test succeeded in this run:
>
> http://www.mobsol.be/buildbot/win32-xp%20VS2005/builds/1915/step-Test%20fsfs%2Bra_local/0
>
> PASS: fs-test.exe 6: check that transaction names are not reused
It's very interesting why this test isn't falling.
> In this older run from r25511, after the txn commits, but before the iconv
> change, tests #22 and #28 were passing:
>
> http://www.mobsol.be/buildbot/win32-xp%20VS2005/builds/1799/step-Test%20fsfs%2Bra_local/0
>
> There's a check I added in trivial_transaction() to validate that transaction
> names contain only the ASCII characters, so the txn names is only ASCII.
>
> How does removing iconv() change any of these things?
Because Subversion became much faster using native Windows iconv.
Tests will pass if you add breakpoints in these tests.
>
> So the longer transaction names may be causing a problem, but why would removing
> iconv support trigger that when the longer names did previously work? Backing
> out the transaction changes will result in shorter transaction names and may
> resolve this issue, but I'm thinking it may not be the real fix for the issue.
Because fsfs doesn't like same transaction name only between revision.
Before you commit it was impossible because transaction name had
revision number.
>
> I suggest that the issue with the Windows clock is a different issue than this
> one and we should see why removing iconv broke the tests.
>
--
Ivan Zhakov
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Fri Jul 27 01:38:49 2007