history-reordering extension merits

[Peter Arrenbrecht]

On Wed, Jan 28, 2009 at 3:11 AM, Matt Mackall <mpm at selenic.com> wrote:
> On Tue, 2009-01-27 at 18:03 -0800, marlborobell wrote:
>>
>> Benoit Boissinot wrote:
>> >
>> > For the record, the manifest of the kernel repo goes from
>> > -rw-r--r--  1 tonfa tonfa 342M 2008-10-23 18:26 00manifest.d.old
>> > to
>> > -rw-r--r--  3 tonfa tonfa 150M 2009-01-27 12:32 00manifest.d
>> >
>> > which is quite impressive (data directory is ~280M big).
>> >
>>
>> Oh. My. Goodness. I just ran this on our repo (with a 1.2G data directory).
>> Our manifest went down from 2GB to 28MB. Yes, by two orders of magnitude.
>> (It took about 40 minutes to run.)
>>
>> I think we have a counterpart to git pack...
>
> No, we just have a known weakness in revlog with regard to heavy
> interleaving. A proper fix (like parent delta) wouldn't need this and
> would probably give even better results.
>
> Also, please note that it is not at all certain your repo will operate
> correctly in all circumstances after rewriting your manifest like this.
> In particular, it may be possible for pull to get confused (the
> "crossing linkrev" issue). So handle with care.

Can you elaborate on the "crossing linkrev" issue? I am worried that
this might affect my approach to shallow clones, too.

-parren