<p><br>
On May 6, 2011 1:55 PM, "Sune Foldager" <<a href="mailto:cryo@cyanite.org">cryo@cyanite.org</a>> wrote:<br>
><br>
> On Fri, May 06, 2011 at 01:43:57 +0200, Benoit Boissinot wrote:<br>
><br>
>> Sune, this makes me think about something we might not have emphasized enough:<br>
>> having an unconsistent deltabase in a revlog (e.g. mixed parent and<br>
>> tip, or even random things when using generaldelta) makes it really<br>
>> hard to generate a bundle suitable for older clients.<br>
><br>
><br>
> Well not hard, just slow, right? You'll end up generating full revisions<br>
> a lot, and calculating their deltas. But that's of course not so good.<br>
></p>
<p>It is really bad, reusing the delta helps a lot. Computing a bit more delta means 10 min instead of 2min generation time on the server. We really don't want that.<br>
><br>
>> This has been one of the big problem for parentdelta (which I hoped<br>
>> could solve by reordering the nodes when generating the bundle, but I<br>
>> need to test that more).<br>
>> I think there's more info on the parentdelta wiki page.<br>
><br>
><br>
> Yeah, reordering might help a lot. Either revlog.group or the bundler<br>
> could do some kinda deltaparent-aware toposort on the revs before the<br>
> actual bundling process. I think that could work.<br>
><br>
> -- <br>
> Sune<br>
</p>