Hi!

I'm a developer working on the on-premise version of Bitbucket at Atlassian.

We noticed pushes to Bitbucket got much slower after upgrading Git on the server side from 2.50 to 2.54. I traced the slow part to the connectivity check that receive-pack runs:

git rev-list --objects --stdin --not --exclude-hidden=receive --all \ --quiet --alternate-refs --progress=Checking connectivity

`strace` shows that after a push, 2.54 does a failing open() of of numerous loose objects -- once in the quarantine (incoming) directory and once in the main object store -- before finding it in a pack:

openat(".../objects/tmp_objdir-incoming-XXXX/ed/58..", O_RDONLY) = ENOENT openat(".../objects/ed/58..", O_RDONLY) = ENOENT

2.50 does not do this. In most customer deployments of Bitbucket, the Git data lives on an NFS share. The extra latency on NFS makes this process of checking for non-existent loose objects take too long, the push essentially hangs at the "Checking connectivity" step.

I believe this new behavior was introduced in the recent object database rework. After using bisect, I belive the problem can be traced back to commit 8384cbcb4c.

I don't know the codebase well, but from what I can see is that the order in which objects are looked up in object databases changed.

Assuming there are two object databases configured (Main repo, and the quarantine directory, for example), the lookup order used to be:

1. _quarantine dir_ packs
2. _main dir_ packs
3. _quarantine dir_ loose objects
4. _main dir_ loose objects

With Git 2.54, the order appears to have changed to:

1. _quarantine dir_ packs
2. _quarantine dir_ loose objects
3. _main dir_ packs
4. _main dir_ loose objects

In my testing, within a well-packed repo, Git 2.50 actually never performed a loose object lookup.

The current design seems to iterate over the configured object databases and perform a pack and loose object lookup for each.

Is there a way to avoid these costly loose object lookups?

On Tue, Jul 21, 2026 at 03:17:23PM +1200, Wolfgang Kritzinger wrote:

Interesting. Here's a smaller reproduction recipe that shows the issue:

  # clone of git.git, or any other non-trivial repo; it should be mostly
  # packed
  src=/path/to/git

  git init empty
  export GIT_ALTERNATE_OBJECT_DIRECTORIES=$src/.git/objects
  strace -fe openat \
    git -C empty rev-list --objects $(git -C $src rev-parse HEAD) >/dev/null

In v2.50, we see almost no loose object open calls, because we check the pack first. But in v2.54, we see tons of them.

Yeah, I can imagine. But even on a fast filesystem, we definitely want to avoid all of those syscalls. Replacing "strace" above with a timing harness, even on a system with fast syscalls and a warm cache, the v2.54 version is ~12% slower.

Hmm, my bisect ended up at a593373b09 (packfile: refactor `find_pack_entry()` to work on the packfile store, 2026-01-09), which is nearby. I'm not sure if it might depend on other factors (e.g., presence of commit graphs, midx, etc) or my reproduction is not exactly like yours, or if one of us messed up bisection.

+cc Patrick as the author of both commits.

Yeah, and that type of regression makes sense for what a593373b09 was trying to do. But I think the v2.54 behavior is wrong. We should check all packs before any loose objects.

I'm not sure of the correct fix. This is working against the whole "odb sources are independent and abstract" refactoring that a593373b09 was going for. But I think it's an important optimization. I guess the abstract version would be that each source has "fast" and "slow" lookups or something like that, and we check all fast ones before slow ones. But that is pretty gross.

I'll leave it to Patrick to ponder further. I haven't really been paying a lot of attention to the odb refactoring.

-Peff

On Mon, Jul 20, 2026 at 11:57:33PM -0400, Jeff King wrote:

I think that both bisection results are equally valid for different reasons.

Before 8384cbcb4c, 'find_pack_entry()' did

    packfile_store_prepare(r->objects->sources->packfiles);

, then tried each of the stores in order to first see if (1) a MIDX was available to locate the object in some pack, or (2) failing that, if there exists some non-MIDX'd pack which could do the same.

Worth noting is that 'packfile_store_prepare()' effectively did:

    for (s = store->source->odb->sources; s; s = s->next) {
        prepare_multi_pack_index_one(s);
        prepare_packed_git_one(s);
    }

Thus preparing the first store also prepared every alternate store, enabling 'find_pack_entry()' to search through all store's MIDX and pack lists/sources.

8384cbcb4c changes this such that 'packfile_store_prepare()' now only prepares its owning source:

    prepare_multi_pack_index_one(store->source);
    prepare_packed_git_one(store->source);

, which is reasonable, but 'find_pack_entry()' still loops over all sources starting from 'r->objects->sources' and calls the function 'packfile_store_prepare()'. But! It calls that function over the same argument each time, like so:

    for (source = r->objects->sources; source; source = source->next) {
        packfile_store_prepare(r->objects->sources->packfiles);
        if (source->midx && fill_midx_entry(source->midx, oid, e))
            return 1;
    }

So we never prepare the packfile store from other sources!

In Wolfgang's case, if we have an quarantine store followed by the main object store, our lookup order will be:

 1. prepare the quarantine object store
 2. search packs in the quarantine object store
 3. search packs in the main object store (which will fail, since this
    list is guaranteed to be empty since we never called
    'packfile_store_prepare()')
 4. search loose objects in the quarantine object store
 5. search loose objects in the main object store
 6. haven't found anything, so we must reprepare
 7. search packs in the main object store, which will now succeed, as
    the previous reprepare called 'packfile_store_prepare()' on the main
    object store's packfile source.

Commit a593373b09 changes things, since it makes 'find_pack_entry()' no longer operate over the entire repository, but over a single store. Before searching that store, it prepares it, like so:

    static int find_pack_entry(struct packfile_store *store,
                               const struct object_id *oid,
                               sturct pack_entry *e)
    {
        struct packfile_list_entry *l;

        packfile_store_prepare(store);
        if (store->source->midx && fill_midx_entry(...))
            return 1;

        for (l = store->packs.head; l; l = l->next) {
            struct packed_git *p = l->pack;
            if (!p->multi_pack_index && fill_pack_entry(oid, e, p)) {
                /* ... */
                return 1;
            }
        }

        return 0;
    }

So commit a593373b09 indeed squashes the bug introduced by 8384cbcb4c, and when lookup reaches the main store, it prepares the main store correctly.

But a593373b09 also changes the lookup order, because the caller in 'do_oid_object_info_extended()` already loops over sources!

    static int do_oid_object_info_extended(struct object_database *odb,
                                           const struct object_id *oid,
                                           struct object_info *oi, unsigned flags)
    {
        /* replace objects, cached lookups, etc., ... */

        odb_prepare_alterantes(odb);

        while (1) {
            struct odb_source *source;

            for (source = odb->sources; source; source = source->next) {
                if (!packfile_store_read_object_info(source->packfiles,
                                                     real, oi, flags) ||
                    !odb_source_loose_read_object_info(source, real, oi,
                                                       flags))
                    return 0;
            }
        }
    }

Before a593373b09, that call to 'packfile_store_read_object_info()' looped over all sources, since it still called 'find_pack_entry()'.

In other words, prior to a593373b09, the lookup proceeded like so:

 1. search packfiles in quarantine
 2. search packfiles in the main object store
 3. search loose objects in quarantine
 4. search loose objects in the main object store

But a593373b09 changes that to instead proceed store-by-store, as follows:

 1. search packfiles in quarantine
 2. search loose objects in quarantine
 3. search packfiles in the main object store
 4. search loose objects in the main object store

So even with all object sources prepared, every object found in a later source pays a failed loose object lookup in an earlier one, which I believe matches what Peff strace'd above.

So both bisections make sense. If the later store has not been prepared yet, commit 8384cbcb4c is where Git first fails to see its packs and falls through to loose object checks. If the stores are already prepared, that problem does not show up, and a593373b09 is where Git first starts checking loose objects in an earlier source before looking in a later source's packs.

I think that something like the following (untested) would fix the immediate issue:

--- 8< ---

Jeff King <peff@peff.net> writes:

I think checking the fast sources before the slow ones is probably the best we can do if we want to retain the 'each odb source is an opaque object' abstraction.

Stepping back a bit, the 'rev-list' command used for the connectivity check is curious in multiple aspects.

 * On the surface, it looks as if the caller wants an enumeration of
   all objects that appear in the range.  However, the caller is not
   interested in the actual list of objects.  Instead, they are
   interested only in a single bit: whether the traversal succeeds
   or dies due to a missing object.  This is because the traversal
   determines whether we need to fetch, or whether we are already up
   to date, to decide whether the proposed 'fetch' is a no-op.  The
   positive ends of the traversal represent what we are about to
   fetch; if we already have all the objects needed to reach those
   tips in our repository, we can do without actually downloading
   anything [*].

 * A false positive answer to the question "does the traversal die
   due to a missing object?" does not affect correctness, as this is
   merely an optimization to save downloads (though a false negative
   is unacceptable).

Given this non-standard use of the command, we can pass application-specific cues (such as "we are doing this traversal for a connectivity check") down to the machinery as a hint to help it optimize its operation, and I suspect that such a hint might have value.

For example, we could enumerate all loose objects in the loose object store using 256 opendir() and readdir() calls for about 10,000 files (since once you have more than 6,700 loose objects, auto-gc would pack them) and store them in an in-core table [**]. This would enable us to say "the object with that name does not exist here" without running lstat() at all. I wonder how many lstat() calls we would need to save for such a scheme to pay off.

There may be other highly application-specific optimization opportunities, as utilizing revision traversal for connectivity checking has peculiar correctness requirements that differ from the normal use of the API.

[Footnote]

 * It follows that in a lazily cloned repository with promisor
   remotes, the traversal could download everything needed as it
   goes, only to conclude: "No need for the main fetch; we have
   everything we need."  I would expect this to be a fairly slow
   process that defeats the entire reason we have this connectivity
   check up front as an optimization.  While I have not checked, I
   believe the actual code prevents this either by skipping the
   connectivity check altogether, or by instructing the connectivity
   checker to treat promised (but not immediately available) objects
   as missing and abort.  But my point is that theoretically one
   does not even need 'git fetch' in a lazily cloned repository.  It
   is sufficient to use 'git ls-remote' to determine the tips of
   remote refs, and run 'rev-list' to fill the range.

** If in-core memory pressure is a concern, we could use a Bloom
   filter, as we only need to know "the object is definitely not
   here" and can tolerate "that object might be here, but we are not
   certain."

On Tue, Jul 21, 2026 at 07:40:02AM -0700, Junio C Hamano wrote:

Seeing that this is about the `tmp_objdir` case: one of the things that Justin and I wanted to work on anyway is that we want to stop modifying the list of sources during transactions in the first place. It always felt kind of gross that we're modifying the sources when creating a transaction, as the only reason that we do this for is so that the writes actually go to the temporary object directory instead of to the primary object source. And that doesn't make a lot of sense to begin with.

The alternative to this would be to instead have logic in functions like `odb_write()` that checks whether we have an active transaction or not. If so, the write would go into the transaction directly instead of going into the primary source, and consequently we wouldn't even have to modify the list of sources at all.

This shouldn't create too much of a problem, as we typically don't intend to even read objects that we've written into the transaction immediately. It would avoid that we try to read objects from the temporary object directory. And it would also allow us to eventually move all the logic to write objects into the transactions exclusively.

I'm currently out of office though, and will be on vacation next week. I'll explore this area a bit more though once I'm back in office in two weeks.

Patrick

Patrick Steinhardt <ps@pks.im> writes:

I suspect several of those 'transactions' are actually misspelt 'temporary directories', but I catch your drift. That said, a redesign like that feels more or less independent of the fix for our immediate performance regression.

After all, didn't Peff show us a case where no odb sources were being flipped in the middle? Simply setting up one object store to borrow from another via the alternates mechanism demonstrated that checking packs across all object stores before hunting for loose objects in any of them makes a world of difference.

Understood.  Bon voyage and have fun!

On Tue, Jul 21, 2026 at 07:40:02AM -0700, Junio C Hamano wrote:

Yeah, I think there may be some interesting opportunities for optimization in check_connected(). But as you noted, it is sometimes used for fetch asking "do we probably have all of these objects" but also for strict connectivity checks for incoming objects. The quarantine area triggers only for the latter (in this case receive-pack), so it would not really help here.

IOW, I'd consider it a mostly orthogonal possible direction for optimization (but still a potentially interesting one).

-Peff

On Wed, Jul 22, 2026 at 01:49:02PM +0200, Patrick Steinhardt wrote:

Yes, the swapping of the "regular" and "alternate" odbs for the quarantine transaction is kind of hacky. But I don't think this is something you can solve just via the odb API. The notion of which sources to read/write from crosses process boundaries. In particular:

  1. We write using a separate index-pack process. It has to be told to
     write into the transaction area, not the regular odb.

  2. After receiving objects, we _do_ read them in order to do quality
     checks before admitting them to the repository. The connectivity
     check discussed here is one example. That happens in a separate
     rev-list process, though it in theory could be moved in-process.

     But we also run user-specified hooks, which may run arbitrary Git
     commands. Those hooks need to be given an environment where they
     can transparently read from both the quarantine area and the
     original odb.

So you'll have to communicate between processes both "write here, not there" and "look at both here and there to read objects". And I suspect the result is going to look a lot like the alternates juggling we are doing today.

-Peff

On Wed, Jul 22, 2026 at 08:49:50AM -0700, Junio C Hamano wrote:

Yeah, exactly. This is really a regression in alternates performance, but it just so happens that the quarantine system is built on top of alternates so we noticed it there.

I'd expect "clone -s / --reference" to have similar problems, and also for sites like GitHub and GitLab that make heavy use of alternates for object sharing between forks. And those would pay the penalty on just about every operation (because we'd expect the alternate to be holding most of the objects in those cases).

-Peff

On Thu, Jul 23, 2026 at 06:49:43AM -0400, Jeff King wrote:

You're right, I also realized that after sending my mail.

One other angle that Justin and I have been discussing (we were at an offsite together over the last couple days) was that we can do a small course correction: instead of handling alternates on the ODB level, we may be able to start treating alternates as an implementation detail of it. So both the handling of alternates, but also the handling of the GIT_OBJECT_DIRECTORY and GIT_ALTERNATE_OBJECT_DIRECTORIES environment variables would be moved into the "files" backend itself.

This would solve a bunch of smaller issues that we're currently grappling with where some of the concepts in Git really want to operate across all of the alternates:

  - The OBJECT_INFO_SECOND_READ flag can be dropped as it becomes an
    implementation detail.

  - We can fix the performance regression because we can now easily
    reorder access to read via packfiles first across all sub-sources.

  - Commit graphs and bitmap really are a singleton, so loading them via
    multiple sources is awkward.

  - The object storage extension that I've written got quite a bit ugly
    as it wasn't quite clear where exactly to draw the line. Especially
    hadnling the environment variables mentioned above into the "files"
    backend removes one point of friction I encountered.

  - Object database maintenance needs to be aware of the other non-local
    sources.

Also, doing that change isn't as bad as it may sound at first. We'd still retain the whole `struct odb_source` list because we want to have them for submodule sources. Furthermore, alternates aren't required for isolation either as we currently use them via the temporary object directory. An alternative implementation may use a completely separate mechanism to achieve write isolation, which is also why we have made the environment variables pluggable that the `struct odb_transaction` ends up passing to the child process.

I think overall this could simplify some of the design, and it makes a bunch of issues that I have been struggling with go away. The devil may be in the details of course, but I think transitioning towards this should be doable.

So I'll work towards that goal. I've got a patch series already that removes all of our calls to `odb_prepare_alternates()` as a first step, but I'll only send that in two weeks once I'm back in office. I'll then have a look at how bad the subsequent steps would be.

Thanks!

Patrick

On Fri, Jul 24, 2026 at 05:46:58AM +0200, Patrick Steinhardt wrote:

That seems reasonable to me. In theory it could lose some flexibility if some code really wanted to represent alternates as abstract sources, but I can't think of why you'd want to do so. Traditionally we did not even really have separate sources at all, and the point of adding them was not so much to have arbitrary combinations of sources as to abstract the details.

I guess somebody could want to use a local alternates object-dir along with their sql database of objects or whatever. But even then, it seems like the alternate could be added as its own files-backend odb source.

Yeah, that all sounds like a positive direction. Thanks!

-Peff

On 26/07/24 05:46AM, Patrick Steinhardt wrote:

Letting the backend control the ordering would be a nice property.

I completely agree. Having the ODB source be more self-contained with the alternates better fits the shape of commits graphs and alternates IMO.

Once all temporary object directory users are updated to use ODB transactions, we could stop reording source list when starting/ending a transaction. Instead the transaction could be tracked separately internally and during ODB read/writes the transaction could be directly used as needed. This is something I plan to tackle in a future series soon.

I am certainly a fan of this direction. We do lose some flexibility in terms of supporting alternates more generically, but I'm not sure supporting alternates of different source types in the same repo would be something we want anyways in practice due to the additional complexity.

-Justin

On Mon, Jul 27, 2026 at 10:45:46AM -0500, Justin Tobler wrote:

[snip]

True. But if we ever find that we actually want that flexibility we don't paint ourselves into a corner, either, as it is rather trivial to introduce another source type that allows us to mix and match different sources.

Patrick