git keeps recreating packs, exploding backup increments

Hi,

I'm trying to understand git's repacking behaviour, as the observed behaviour doesn't match how I read the documentation or the code.

The problem we see is excessive backup increments for developer directories. The cause is that pack files keep getting regenerated for large repositories.

Users are not running 'git gc' manually, so the assumption is that this is caused by 'git gc --auto' being run implicitly.

 From what I can see in the code, and the documentation, it should only 
pack up objects not already found in existing packs. Or at the very 
least, not the objects found in the largest existing pack.

(at least not until gc.autoPackLimit is hit)

But this isn't happening. Old packs are constantly being replaced by new ones. Despite most of the objects being old and stable.

We tried gc.bigPackThreshold in the hope it would force it to reuse packs better. But all we got instead was duplication. It still creates new packs with everything. It just stopped removing the old ones.

Some guidance would be appreciated. I cannot find anything in the code or documentation that explains the current behaviour.

Regads,

On Wed, Feb 19, 2025 at 5:58 PM Pierre Ossman <ossman@cendio.se> wrote:

Is the repo partially cloned? git-repack will always pack promisor packs even if it's a keep pack. This patch would fix it https://lore.kernel.org/git/2728513.vuYhMxLoTh@mintaka.ncbr.muni.cz/

Thanks

On 20/02/2025 04:03, Han Young wrote:

Yes, the big offender is often partially cloned. So that could be part of it, thanks.

But we're seeing it in other repositories as well. E.g. I have a long-lived TigerVNC repository where the biggest pack file is just one week old. In that case, it's merely 21 MiB, so it's not a practical issue. But it does show that git keeps replacing it.

Anything I/we can do to shed more light on the issue?

Regards,

On Thu, Feb 20, 2025 at 09:26:38AM +0100, Pierre Ossman wrote:

Well, one of the interesting things to learn would be how often you end up updating those repositories. You have discovered "gc.autoPackLimit" already, which determines when exactly Git is going to repack existing packfiles into one, and mentioned that it doesn't seem to help you. But whether it does or doesn't help really depends on how frequently you gain new packfiles in the impacted repositories.

When you have fast-moving repositories and developers fetch several times per day, then it is quite likely that they accumulate multiple new packfiles per day. And thus, it's not all that unexpected that you will have to repack the whole repository rather regularly. If so, this is working as designed. You can tune the parameters for how often Git will do an all-into-one repack, but also have to keep in mind that the more packfiles there are, the less efficient Git will in general be.

That being said, there is an alternative: Git nowadays doesn't use git-gc(1) anymore to perform auto-maintenance, but instead it invokes git-maintenance(1). And that command allows the user to pick what tasks should be performed. By default it uses git-gc(1) under the hood indeed, but you also ask it to not do so and instead use an alternative mechanism to pack your objects.

The alternative would be the "incremental-repack" task. This task does not use git-gc(1) with its incremental/all-into-one repack split, but it instead uses git-multi-pack-index(1). git-maintenance(1) tweaks the `--batch-size` parameter of `git multi-pack-index repack` so that it typically doesn't have to repack the one large packfile, but combines at least two smaller ones. I use a mechanism like that, which I've configured as follows:

    [maintenance "commit-graph"]
        enabled = true
    [maintenance "gc"]
        enabled = false
    [maintenance "incremental-repack"]
        enabled = true
    [maintenance "loose-objects"]
        enabled = true
    [maintenance "pack-refs"]
        enabled = true

I think this strategy still isn't quite optimal, as nowadays we should probably make use of `git repack --geometric` instead of manually computing batch sizes. This would ensure that the packfiles present in the repository form a geometric sequence regarding their size, so you end up repacking the biggest packfile very infrequently. Such a task has not been implemented yet, but it shouldn't be all that hard to do, either.

Patrick

On 21/02/2025 09:16, Patrick Steinhardt wrote:

I don't think the most problematic repo should be moving that fast. But I might be wrong. We've reverted all settings to default, and we'll try to keep an eye on what happens to the pack files to gain more understanding.

Thanks. This is definitely something we can try. We'll observe the system for now, to establish a new baseline. Then we'll try some of these settings and see how it affect things.

Regards,

Following up on this old thread.

I think the entire thing might be a false diagnostic on our part, and git's default behaviour is working just fine for us.

We initially starting looking at this because the backups exploded in size. But git is complex enough that we had a hard time examining things retroactively. Instead, we started monitoring things going forward to find likely causes.

What we did was to keep an eye on how new the files were in .git/objects/pack. And we kept seeing that the large packs were brand spanking new. Hence, the attempts at reconfiguring git, and the thread here.

But it seems we didn't look close enough. Although the timestamps suggest that the packs are constantly being modified, the names and contents actually stay the same.

I've been keeping a closer eye on a couple of active repositories, and we aren't actually seeing any excessive growth in size. But the largest pack files always have a very current modification time.

No idea what caused that initial spike in backup storage. We'll have to revisit that if it happens again.

Regards,