[webkit-dev] svn.webkit.org

[Mark Rowe]

On 2012-03-01, at 04:28, Osztrogonac Csaba <oszi at inf.u-szeged.hu> wrote:

> Mark Rowe írta:
>> On 2012-03-01, at 03:37, Osztrogonac Csaba <oszi at inf.u-szeged.hu> wrote:
>>> After unbanning our network, I'll copy a locally tar-ed WebKit-svn
>>> copy to all bots and then restart them one by one not to overload
>>> svn.webkit.org.
>> You can get a relatively up to date working copy from <http://nightly.webkit.org/files/WebKit-SVN-source.tar.bz2>.
> 
> We always have relatively up-to-date working copy. It is faster then downloading
> from you. ;) (I tried to download this nightly and the ETA is ~3 hours ...)
> Additionally we use svn 1.7 (to make our bots faster, and less IO intensive)
> which has different working copy than svn 1.6 .

I'm surprised that it downloads so slowly for you. Do you have a rough idea how long a fresh "svn checkout" of WebKit trunk takes? I'd expect the tarball to download more quickly than svn can fetch.

>>> And I'm thinking about how can we make buildbots more robust in the future.
>>> My first idea is that we should setup git mirrors for all WebKit port, and
>>> then make buildbots use these mirrors instead of svn.webkit.org. To do it,
>>> we need to modify the master.cfg a little bit.
>> There are two different options that we're investigating to address this "thundering herd" problem that tends to kill SVN after buildbot downtime:
>> 1) Teach build slaves to fetch and unpack <http://nightly.webkit.org/files/WebKit-SVN-source.tar.bz2> rather than using "svn checkout".
> 
> To do it, all slave must use same svn version.

They'd need to use the same version or newer. It's trivial to have the slave upgrade the working copy if necessary after unpacking it.

>> 2) Add additional hardware and load-balance svn.webkit.org across several machines so that the spike in load is distributed.
> 
> It is a good idea. ;)
> 
>>> To migrate the bots simple from svn to git isn't a good solution, because in this case
>>> we'll loose the svn revision number on the waterfall. My idea is that replace the
>>> "self.addStep(CheckOutSource)" to calling a shell/perl/python script which updates
>>> the local copy from the Qt/GTK/Apple/Chromium/... git mirror with the following way:
>>> - git fetch
>>> - git reset --hard `git svn find-rev r<svn-revision-number-got-from-the-master>`
>>> 
>>> I'm going to implement this initial git updater script this week and try to migrate our
>>> bots on build.webkit.sed.hu to use it. If we manage to make it stable, we can make
>>> build.webkit.org slaves to use it too. How does it sound? Any better idea?
>> Pulling from git instead of SVN is certainly worth considering, but I wouldn't be surprised if this simply shifted the performance issue from svn.webkit.org to git.webkit.org.
> 
> No, I don't want to shift the load from svn.webkit.org to git.webkit.org. But make all port
> maintainer to use their own git mirror for the bots instead of the root svn/git.webkit.org.
> Mirroring git.webkit.org locally is the simplest thing in the world. :)

I'd prefer to avoid forcing people to set up their own mirrors if at all possible.

- Mark