Re: [PATCH] midx.c: use `pack-objects --stdin-packs` when repacking

From: Taylor Blau <me@ttaylorr.com>
Date: Sep 20, 2022, 20:35 UTC
Message-ID: <YyokIf/Sd7SYztKQ@nand.local>
In-Reply-To: <YyodQg5diyr/UYK1@coredump.intra.peff.net>

On Tue, Sep 20, 2022 at 04:06:26PM -0400, Jeff King wrote:

Show 20 quoted lines

> On Tue, Sep 20, 2022 at 03:49:07PM -0400, Taylor Blau wrote:
>
> > > Is that true also of "multi-pack-index repack"? I guess it would depend
> > > on how you invoke it. I admit I don't think I've ever used it myself,
> > > since the new "repack --geometric --write-midx" approach matches my
> > > mental model. I'm not sure when you'd actually run the "multi-pack-index
> > > repack" command. But if you did it with --batch-size=0 (the default), I
> > > think we'd end up traversing every object in history.
> >
> > We could probably benefit from it, but only if there is a MIDX bitmap
> > around to begin with. For instance, you could first try and lookup each
> > object you're missing a namehash for and then read its value out of the
> > hashcache extension in the MIDX bitmap (assuming the MIDX bitmap exists,
> > and has a hashcache).
> >
> > But if you don't have a MIDX bitmap, or it has a poor selection of
> > commits, then you're out of luck.
>
> You could also use a pack bitmap if there is one (and it's one of the
> included packs). But yes, if you have neither, it's no help.

Good point. But, yeah, you have to have them to begin with.

Show 21 quoted lines

> Mostly I'm just concerned that this could have an outsized negative
> performance effect if you have a setup like:
>
>   - you have a gigantic repository, say that takes 15 minutes to do a
>     full "rev-list --objects" on (like linux.git with all its forks)
>
>   - most of that is in one big pack, but you acquire new packs
>     occasionally via pushes, etc
>
>   - doing "git repack --geometric" rolls up the new packs, nicely
>     traversing just the new objects
>
>   - doing "git multi-pack-index repack" before your patch is fast-ish.
>     It stuffs all the objects into a new pack. But after your patch, it
>     does that 15-minute traversal.
>
> But I don't know if that's even realistic, because I'm still wondering
> why somebody would run "git multi-pack-index repack" in the first place.
> And if they'd always do so with --batch-size anyway, which would
> mitigate this (because it gives a geometric-ish approach where we leave
> the huge pack untouched).

Yeah, the `--geometric` path(s) don't have this problem, because the big pack will already be covered by either a pack or MIDX bitmap, and we can read out all of the namehash values from there.

But I tend to agree that this is pretty unrealistic, so I'm hopeful that it isn't a huge deal. If it is, though, we can always just "turn off" the traversal parts. (Though I have to imagine that a repository large enough to care about the existence of namehash values probably isn't using `git multi-pack-index repack` anyway).

Show 6 quoted lines

> Yeah, sorting the packs by mtime might be sensible. I know in the final
> midx, we use object order to find the "preferred" pack. And you could
> iterate the objects here, passing along their de-duped pack name. But I
> don't think we have the objects here in that useful order; that is
> really the order of the midx's .rev file, IIRC, and this is probably the
> actual sha1 order.

We already need the sorted order in order to compute the rollup for non-zero `--batch-size` arguments, so using that to construct the pack is just a matter of dragging the sort out of the function to compute the rollup itself (and into `midx_repack()` instead).

Patches incoming... ;-)

Thanks, Taylor