Re: [PATCH v2 09/24] midx: infer preferred pack when not given one

From: Taylor Blau <me@ttaylorr.com>
To: Jeff King <peff@peff.net>
Cc: git@vger.kernel.org, dstolee@microsoft.com, gitster@pobox.com,
	jonathantanmy@google.com
Subject: Re: [PATCH v2 09/24] midx: infer preferred pack when not given one
Date: Wed, 21 Jul 2021 16:16:07 -0400	[thread overview]
Message-ID: <YPiAhw2eP2MOksUF@nand.local> (raw)
In-Reply-To: <YPf4MTDpbvinoIia@coredump.intra.peff.net>

On Wed, Jul 21, 2021 at 06:34:25AM -0400, Jeff King wrote:
> On Mon, Jun 21, 2021 at 06:25:21PM -0400, Taylor Blau wrote:
>
> > In 9218c6a40c (midx: allow marking a pack as preferred, 2021-03-30), the
> > multi-pack index code learned how to select a pack which all duplicate
> > objects are selected from. That is, if an object appears in multiple
> > packs, select the copy in the preferred pack before breaking ties
> > according to the other rules like pack mtime and readdir() order.
> >
> > Not specifying a preferred pack can cause serious problems with
> > multi-pack reachability bitmaps, because these bitmaps rely on having at
> > least one pack from which all duplicates are selected. Not having such a
> > pack causes problems with the pack reuse code (e.g., like assuming that
> > a base object was sent from that pack via reuse when in fact the base
> > was selected from a different pack).
>
> It might be helpful to use a more descriptive name for "pack reuse code"
> here, since it's kind of vague for people who have not been actively
> working on bitmaps.
>
> I don't have a short name for that chunk of code, but maybe:
>
>   ...causes problems with the code in pack-objects to reuse packs
>   verbatim (e.g., that code assumes that a delta object in a chunk of
>   pack sent verbatim will have its base object sent from the same pack).

Thanks; I like what you wrote here.

> >   - The psuedo pack-order (described in
> >     Documentation/technical/bitmap-format.txt) is computed by
> >     midx_pack_order(), and sorts by pack ID and pack offset, with
> >     preferred packs sorting first.
>
> I think the .rev description in pack-format.txt may be a better
> reference here.

Ditto, I changed that, too.

> >   - But! Pack IDs come from incrementing the pack count in
> >     add_pack_to_midx(), which is a callback to
> >     for_each_file_in_pack_dir(), meaning that pack IDs are assigned in
> >     readdir() order.
> >
> > [ ... ]
>
> This explanation is rather confusing, but I'm not sure if we can do much
> better. I followed all of it, because I was there when we found the bug
> that this is fixing. And of course that happened _after_ we implemented
> midx bitmaps and in particular adapted the verbatim reuse stuff in
> pack-objects to make use of it.
>
> I see why you'd want to float the fix up before then, so we don't ever
> have the broken state. But it's hard to understand what bug this is
> fixing, because the bug does not even exist yet at this point in
> the series!
>
> I dunno. Like I said, I was able to follow it, so maybe it is
> sufficient. I'm just not sure others would be able to.

I think that others will follow it, too. But I agree that it is
confusing, since we're fixing a bug that doesn't yet exist. In reality,
I wrote this patch after sending v1, and then reordered its position to
come before the implementation of MIDX bitmaps for that reason.

So in one sense, I prefer it this way because we don't ever introduce
the bug.  But in another sense, it is very jarring to read about an
interaction that has no basis in the code (yet).

I think that the best thing we could do without adding any significant
reordering would be to just call out the situation we're in. I added
this onto the end of the commit message which I think makes things a
little clearer:

    (Note that multi-pack reachability bitmaps have yet to be
    implemented; so in that sense this patch is fixing a bug which does
    not yet exist.  But by having this patch beforehand, we can prevent
    the bug from ever materializing.)

Thanks,
Taylor