Re: [PATCH 0/2] prevent `repack` to unpack and delete promisor objects

From: Jeff King <peff@peff.net>
To: Rafael Silva <rafaeloliveira.cs@gmail.com>
Cc: git@vger.kernel.org, "SZEDER Gábor" <szeder.dev@gmail.com>,
	"Jonathan Tan" <jonathantanmy@google.com>
Subject: Re: [PATCH 0/2] prevent `repack` to unpack and delete promisor objects
Date: Thu, 15 Apr 2021 05:15:04 -0400	[thread overview]
Message-ID: <YHgEGHIgwfobcwDr@coredump.intra.peff.net> (raw)
In-Reply-To: <20210414191403.4387-1-rafaeloliveira.cs@gmail.com>

On Wed, Apr 14, 2021 at 09:14:01PM +0200, Rafael Silva wrote:

> It took me a bit to come up with the test because it seems `repack`
> doesn't offer an option to skip the "deletion of unpacked objects",
> so this series adds a new option to `repack` for skip the
> `git prune-packed` execution thus allowing us to easily inspect the
> unpacked objects before they are removed and simplification of our
> test suite. Furthermore, The test will now test the `repack` code
> path instead of performing the operations by calling
> `pack-objects`.

Thanks for working on this. Overall the patches seem sane, though I
think Jonathan's comments (especially about the confusion in the commit
message of 2/2) are worth addressing.

I have mixed feelings on the "--no-prune-packed" option, just because
it's user-visible and I don't think it's something a normal user would
ever really want.

In the new test (and I think in the old ones you modified, though I
didn't look carefully) the main thing we care about is whether we write
out loose objects. So another solution would be to improve the debug
logging inside pack-objects to tell us more about what it's doing.

The fork of Git we use at GitHub has something similar; when we discard
objects or force them loose, we write their sha1 values to a log file.
This has come in handy for a lot of after-the-fact debugging ("oops,
this repo is corrupted; did we intentionally delete object X?").

I wonder if we could do something similar with the trace2 facility. I
know it can be turned on via config, but I don't know how good the
support is for enabling just one segment of data (and this may generate
a lot of entries, so people using trace2 for telemetry probably wouldn't
want it on).

For the purposes of the tests, though just a normal GIT_TRACE_PACK_DEBUG
would be plenty. I dunno. I don't want to open up a can of worms on
logging that would hold up getting this quite-substantial fix in place.
But once we add --no-prune-packed, it will be hard to take away.

-Peff