PATH WALK III: Add 'git backfill' command

[derrickstolee]

This is based on v3 of ds/path-walk-1 [1] and an earlier version was part of my initial path-walk RFC [2].

[1] https://lore.kernel.org/git/[email protected]/

[2] https://lore.kernel.org/git/[email protected]/

This series adds a new 'git backfill' command that uses the path-walk API to download missing blobs in a blobless partial clone. Users can specify interaction with the sparse-checkout using '--[no-]sparse' but the '--sparse' option is implied by the existence of a sparse-checkout.

The reason to use the path-walk API is to make sure that the missing objects are grouped by a common path, giving a reasonable process for batching requests and expecting the server to compress the resulting packfile nicely together.

I first prototyped this feature in June 2024 as an exploration and created the path-walk algorithm for this purpose. It was only my intuition that led me to believe that batching by path would lead to better packfiles. This has been proven out as a very important feature due to recent investigations to compressing full repositories by doing a better job of grouping objects by path. See the --name-hash-version series [3] or the 'git pack-objects --path-walk' series [4] (currently on hold as it conflicts with the --name-hash-version series).

[3] https://lore.kernel.org/git/[email protected]/

[4] https://lore.kernel.org/git/[email protected]/

This idea can be further demonstrated by the evidence in testing this feature: by downloading objects in small batch sizes, the client can force the server to repack things more efficiently than a full repack.

The example repository I have used in multiple places is the microsoft/fluentui repo [5] as it has many CHANGELOG.md files that cause name hash collisions that make the full repack inefficient.

[5] https://github.com/microsoft/fluentui

If we create a blobless clone of the fluentui repo, then this downloads 105 MB across two packfiles (the commits and trees pack, followed by the blobs needed for an initial checkout). Running 'git backfill --batch-size=' for different sizes leads to some interesting results:

| Batch Size      | Pack Count | Pack Size | Time   |
|-----------------|------------|-----------|--------|
| (Initial clone) | 2          | 105 MB    |        |
| 5K              | 53         | 348 MB    | 2m 26s |
| 10K             | 28         | 365 MB    | 2m 22s |
| 15K             | 19         | 407 MB    | 2m 21s |
| 20K             | 15         | 393 MB    | 2m 28s |
| 25K             | 13         | 417 MB    | 2m 06s |
| 50K             | 8          | 509 MB    | 1m 34s |
| 100K            | 5          | 535 MB    | 1m 56s |
| 250K            | 4          | 698 MB    | 1m 33s |
| 500K            | 3          | 696 MB    | 1m 42s |

The smaller batches cause the server to realize that the existing deltas cannot be reused and it finds better deltas. This takes some extra time for the small batches, but halves the size of the repo. Even in the 500K batch size, we get less data than the 738 MB of a full clone.

Implementing the --sparse feature is best done by augmenting the path-walk API to be aware of a pattern list. This works for both cone and non-cone mode sparse-checkouts.

There are future directions we could take this command, especially to run the command with a user-specified pathspec. The tricky case for that additional feature is trying to make the path-walk more efficient by skipping tree paths that would not lead to a match of the pathspec. It would likely need optimization in a small subset of pathspec features (such as prefix matches) to work as efficiently as possible. I did prototype a version that puts the pathspec match in the callback function within builtin/backfill.c, but I found that uninspiring and unnecessary for now.

Thanks, -Stolee

cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]
cc: [email protected]

[derrickstolee]

/submit

[gitgitgadget]

Submitted as [email protected]

To fetch this version into FETCH_HEAD:

git fetch https://github.com/gitgitgadget/git/ pr-1820/derrickstolee/backfill-upstream-v1

To fetch this version to local tag pr-1820/derrickstolee/backfill-upstream-v1:

git fetch --no-tags https://github.com/gitgitgadget/git/ tag pr-1820/derrickstolee/backfill-upstream-v1

[gitgitgadget]

This patch series was integrated into seen via git@e1db44d.

[gitgitgadget]

This patch series was integrated into seen via git@e1dd1e4.

[gitgitgadget]

On the Git mailing list, Junio C Hamano wrote (reply to this):

It seems that the result of calling init_revisions() from backfill
is leaked?

https://github.com/git/git/actions/runs/12218430154/job/34083929479 

I did not dig further but the below is from my local leaksanitizer
run.

Thanks.

=================================================================
==git==182342==ERROR: LeakSanitizer: detected memory leaks

Direct leak of 576 byte(s) in 1 object(s) allocated from:
    #0 0x55d4b0e42915 in __interceptor_realloc (git+0x84915) (BuildId: c861e65ec43b0a3ef46b9555a81d6ddfc2358e8e)
    #1 0x55d4b119ce6d in xrealloc wrapper.c:140:8
    #2 0x55d4b11204d4 in add_rev_cmdline revision.c:1563:2
    #3 0x55d4b11187a1 in handle_revision_arg_1 revision.c:2263:2
    #4 0x55d4b1118398 in handle_revision_arg revision.c:2275:12
    #5 0x55d4b0e5233b in do_backfill builtin/backfill.c:100:2
    #6 0x55d4b0e52253 in cmd_backfill builtin/backfill.c:146:9
    #7 0x55d4b0e46b80 in run_builtin git.c:480:11
    #8 0x55d4b0e45502 in handle_builtin git.c:741:9
    #9 0x55d4b0e4649c in run_argv git.c:808:4
    #10 0x55d4b0e45274 in cmd_main git.c:948:19
    #11 0x55d4b0f6da8a in main common-main.c:9:11
    #12 0x7ff25f97ac89 in __libc_start_call_main csu/../sysdeps/nptl/libc_start_call_main.h:58:16
    #13 0x7ff25f97ad44 in __libc_start_main csu/../csu/libc-start.c:360:3
    #14 0x55d4b0e171e0 in _start (git+0x591e0) (BuildId: c861e65ec43b0a3ef46b9555a81d6ddfc2358e8e)

DEDUP_TOKEN: __interceptor_realloc--xrealloc--add_rev_cmdline--handle_revision_arg_1--handle_revision_arg--do_backfill--cmd_backfill--run_builtin--handle_builtin--run_argv--cmd_main--main--__libc_start_call_main--__libc_start_main--_start
Indirect leak of 5 byte(s) in 1 object(s) allocated from:
    #0 0x55d4b0e424b6 in __interceptor_malloc (git+0x844b6) (BuildId: c861e65ec43b0a3ef46b9555a81d6ddfc2358e8e)
    #1 0x7ff25f9f34f9 in strdup string/strdup.c:42:15
    #2 0x55d4b119cb14 in xstrdup wrapper.c:43:14
    #3 0x55d4b1120506 in add_rev_cmdline revision.c:1565:23
    #4 0x55d4b11187a1 in handle_revision_arg_1 revision.c:2263:2
    #5 0x55d4b1118398 in handle_revision_arg revision.c:2275:12
    #6 0x55d4b0e5233b in do_backfill builtin/backfill.c:100:2
    #7 0x55d4b0e52253 in cmd_backfill builtin/backfill.c:146:9
    #8 0x55d4b0e46b80 in run_builtin git.c:480:11
    #9 0x55d4b0e45502 in handle_builtin git.c:741:9
    #10 0x55d4b0e4649c in run_argv git.c:808:4
    #11 0x55d4b0e45274 in cmd_main git.c:948:19
    #12 0x55d4b0f6da8a in main common-main.c:9:11
    #13 0x7ff25f97ac89 in __libc_start_call_main csu/../sysdeps/nptl/libc_start_call_main.h:58:16
    #14 0x7ff25f97ad44 in __libc_start_main csu/../csu/libc-start.c:360:3
    #15 0x55d4b0e171e0 in _start (git+0x591e0) (BuildId: c861e65ec43b0a3ef46b9555a81d6ddfc2358e8e)

DEDUP_TOKEN: __interceptor_malloc--strdup--xstrdup--add_rev_cmdline--handle_revision_arg_1--handle_revision_arg--do_backfill--cmd_backfill--run_builtin--handle_builtin--run_argv--cmd_main--main--__libc_start_call_main--__libc_start_main--_start
SUMMARY: LeakSanitizer: 581 byte(s) leaked in 2 allocation(s).
Our logs revealed a memory leak...
++ rmdir /home/gitster/w/git.git/t/test-results/t5620-backfill.leak
++ :
++ exit 134
++ eval_ret=134

[gitgitgadget]

On the Git mailing list, Junio C Hamano wrote (reply to this):

Junio C Hamano <[email protected]> writes:

> It seems that the result of calling init_revisions() from backfill
> is leaked?
>
> https://github.com/git/git/actions/runs/12218430154/job/34083929479 
>
> I did not dig further but the below is from my local leaksanitizer
> run.

The attached patch seems to plug the leaks observed by your backfill
tests.  If you agree with the implementation of the change, you are
welcome to squash it in.  I may be missing a better mechanism in the
path-walk API that I could use to plug the leaks, in which case, of
course a fix using such a better mechanism is very much welcomed.

A few things I noticed about path-walk API during my poking are:

 - struct path_walk_info has PATH_WALK_INFO_INIT() macro, but not a
   matching path_walk_info_init() helper function to help initialize
   any heap-allocated instances.  In general, anything that requires
   _INIT() macro by definition wants to be initialized with more
   than nul-initialized (otherwise, it would be fine to leave it
   uninitialized in the BSS, clear with = {0} initialization on the
   stack), so lack of matching path_walk_info_init() raises an
   eyebrow.

 - struct path_walk_info seems to lack a destructor.  In general,
   anything complex enough to require initializer should have one.

 - lack of destructor for path_walk_info requires users to release
   resources held by "struct pattern_list" instance at pwi.pl; if we
   had a destructor, we wouldn't have 2 of the three leaks we had to
   fix here.

Thanks.


diff --git c/builtin/backfill.c w/builtin/backfill.c
index 225764f17e..1cf2df9303 100644
--- c/builtin/backfill.c
+++ w/builtin/backfill.c
@@ -92,8 +92,11 @@ static int do_backfill(struct backfill_context *ctx)
 
 	if (ctx->sparse) {
 		CALLOC_ARRAY(info.pl, 1);
-		if (get_sparse_checkout_patterns(info.pl))
+		if (get_sparse_checkout_patterns(info.pl)) {
+			clear_pattern_list(info.pl);
+			free(info.pl);
 			return error(_("problem loading sparse-checkout"));
+		}
 	}
 
 	repo_init_revisions(ctx->repo, &revs, "");
@@ -113,6 +116,11 @@ static int do_backfill(struct backfill_context *ctx)
 		download_batch(ctx);
 
 	clear_backfill_context(ctx);
+	release_revisions(&revs);
+	if (info.pl) {
+		clear_pattern_list(info.pl);
+		free(info.pl);
+	}
 	return ret;
 }
 

[gitgitgadget]

This patch series was integrated into seen via git@ef760f5.

[gitgitgadget]

This patch series was integrated into seen via git@6eef8e2.

[gitgitgadget]

This patch series was integrated into seen via git@7668c77.

[gitgitgadget]

This patch series was integrated into seen via git@8a5ce5d.

[gitgitgadget]

This patch series was integrated into seen via git@28b8993.

[gitgitgadget]

This patch series was integrated into seen via git@3c075bc.

[gitgitgadget]

This patch series was integrated into seen via git@015ab1a.

[gitgitgadget]

This patch series was integrated into seen via git@8def160.

[gitgitgadget]

On the Git mailing list, Patrick Steinhardt wrote (reply to this):

On Fri, Dec 06, 2024 at 08:07:16PM +0000, Derrick Stolee via GitGitGadget wrote:
> From: Derrick Stolee <[email protected]>
> 
> Users may want to specify a minimum batch size for their needs. This is only
> a minimum: the path-walk API provides a list of OIDs that correspond to the
> same path, and thus it is optimal to allow delta compression across those
> objects in a single server request.

Okay, here you explicitly say that this is a minimum batch size, so this
is by design and with a proper reason. Good.

> We could consider limiting the request to have a maximum batch size in the
> future. For now, we let the path-walk API batches determine the
> boundaries.

Should we maybe rename `--batch-size` to `--min-batch-size` so that it
does not become awkward if we ever want to have a maximum batch size, as
well? Also helps to set expectations with the user.

[snip]
> Based on these experiments, a batch size of 50,000 was chosen as the
> default value.

Thanks for all the data, this is really helpful!

> diff --git a/Documentation/git-backfill.txt b/Documentation/git-backfill.txt
> index 0e10f066fef..9b0bae04e9d 100644
> --- a/Documentation/git-backfill.txt
> +++ b/Documentation/git-backfill.txt
> @@ -38,6 +38,14 @@ delta compression in the packfile sent by the server.
>  By default, `git backfill` downloads all blobs reachable from the `HEAD`
>  commit. This set can be restricted or expanded using various options.
>  
> +OPTIONS
> +-------
> +
> +--batch-size=<n>::
> +	Specify a minimum size for a batch of missing objects to request
> +	from the server. This size may be exceeded by the last set of
> +	blobs seen at a given path. Default batch size is 16,000.

This is stale: s/16,000/50,000/

> diff --git a/builtin/backfill.c b/builtin/backfill.c
> index e5f2000d5e0..127333daef8 100644
> --- a/builtin/backfill.c
> +++ b/builtin/backfill.c
> @@ -112,6 +112,8 @@ int cmd_backfill(int argc, const char **argv, const char *prefix,
> struct reposit
>                 .batch_size = 50000,
>         };
>         struct option options[] = {
> +               OPT_INTEGER(0, "batch-size", &ctx.batch_size,
> +                           N_("Minimun number of objects to request at a time")),

s/Minimun/Minimum

Patrick

[gitgitgadget]

On the Git mailing list, Derrick Stolee wrote (reply to this):

On 12/16/24 3:01 AM, Patrick Steinhardt wrote:
> On Fri, Dec 06, 2024 at 08:07:16PM +0000, Derrick Stolee via GitGitGadget wrote:
>> From: Derrick Stolee <[email protected]>
>>
>> Users may want to specify a minimum batch size for their needs. This is only
>> a minimum: the path-walk API provides a list of OIDs that correspond to the
>> same path, and thus it is optimal to allow delta compression across those
>> objects in a single server request.
>
> Okay, here you explicitly say that this is a minimum batch size, so this
> is by design and with a proper reason. Good.
>
>> We could consider limiting the request to have a maximum batch size in the
>> future. For now, we let the path-walk API batches determine the
>> boundaries.
>
> Should we maybe rename `--batch-size` to `--min-batch-size` so that it
> does not become awkward if we ever want to have a maximum batch size, as
> well? Also helps to set expectations with the user.

Good idea. Will do.

>> diff --git a/Documentation/git-backfill.txt b/Documentation/git-backfill.txt
>> index 0e10f066fef..9b0bae04e9d 100644
>> --- a/Documentation/git-backfill.txt
>> +++ b/Documentation/git-backfill.txt
>> @@ -38,6 +38,14 @@ delta compression in the packfile sent by the server.
>>   By default, `git backfill` downloads all blobs reachable from the `HEAD`
>>   commit. This set can be restricted or expanded using various options.
>>
>> +OPTIONS
>> +-------
>> +
>> +--batch-size=<n>::
>> +	Specify a minimum size for a batch of missing objects to request
>> +	from the server. This size may be exceeded by the last set of
>> +	blobs seen at a given path. Default batch size is 16,000.
>
> This is stale: s/16,000/50,000/

Thanks!

>> diff --git a/builtin/backfill.c b/builtin/backfill.c
>> index e5f2000d5e0..127333daef8 100644
>> --- a/builtin/backfill.c
>> +++ b/builtin/backfill.c
>> @@ -112,6 +112,8 @@ int cmd_backfill(int argc, const char **argv, const char *prefix,
>> struct reposit
>>                  .batch_size = 50000,
>>          };
>>          struct option options[] = {
>> +               OPT_INTEGER(0, "batch-size", &ctx.batch_size,
>> +                           N_("Minimun number of objects to request at a time")),
>
> s/Minimun/Minimum

Thanks for the careful eye for detail.

-Stolee

[gitgitgadget]

On the Git mailing list, Patrick Steinhardt wrote (reply to this):

On Fri, Dec 06, 2024 at 08:07:15PM +0000, Derrick Stolee via GitGitGadget wrote:
> diff --git a/Documentation/git-backfill.txt b/Documentation/git-backfill.txt
> index 640144187d3..0e10f066fef 100644
> --- a/Documentation/git-backfill.txt
> +++ b/Documentation/git-backfill.txt
> @@ -14,6 +14,30 @@ SYNOPSIS
>  DESCRIPTION
>  -----------
>  
> +Blobless partial clones are created using `git clone --filter=blob:none`
> +and then configure the local repository such that the Git client avoids
> +downloading blob objects unless they are required for a local operation.
> +This initially means that the clone and later fetches download reachable
> +commits and trees but no blobs. Later operations that change the `HEAD`
> +pointer, such as `git checkout` or `git merge`, may need to download
> +missing blobs in order to complete their operation.

Okay.

> +In the worst cases, commands that compute blob diffs, such as `git blame`,
> +become very slow as they download the missing blobs in single-blob
> +requests to satisfy the missing object as the Git command needs it. This
> +leads to multiple download requests and no ability for the Git server to
> +provide delta compression across those objects.
> +
> +The `git backfill` command provides a way for the user to request that
> +Git downloads the missing blobs (with optional filters) such that the
> +missing blobs representing historical versions of files can be downloaded
> +in batches. The `backfill` command attempts to optimize the request by
> +grouping blobs that appear at the same path, hopefully leading to good
> +delta compression in the packfile sent by the server.

Hm. So we're asking the user to fix a usability issue of git-blame(1),
don't we? Ideally, git-blame(1) itself should know to transparently
batch the blobs it requires to compute its output, shouldn't it? That
usecase alone doesn't yet convince me that git-backfill(1) is a good
idea as I'd think we should rather fix the underlying issue.

So are there other usecases for git-backfill(1)? I can imagine that it
might be helpful in the context of scripts that know they'll operate on
a bunch of blobs.

> diff --git a/builtin/backfill.c b/builtin/backfill.c
> index 38e6aaeaa03..e5f2000d5e0 100644
> --- a/builtin/backfill.c
> +++ b/builtin/backfill.c
> @@ -1,16 +1,116 @@
>  #include "builtin.h"
> +#include "git-compat-util.h"
>  #include "config.h"
>  #include "parse-options.h"
>  #include "repository.h"
> +#include "commit.h"
> +#include "hex.h"
> +#include "tree.h"
> +#include "tree-walk.h"
>  #include "object.h"
> +#include "object-store-ll.h"
> +#include "oid-array.h"
> +#include "oidset.h"
> +#include "promisor-remote.h"
> +#include "strmap.h"
> +#include "string-list.h"
> +#include "revision.h"
> +#include "trace2.h"
> +#include "progress.h"
> +#include "packfile.h"
> +#include "path-walk.h"
>  
>  static const char * const builtin_backfill_usage[] = {
>  	N_("git backfill [<options>]"),
>  	NULL
>  };
>  
> +struct backfill_context {
> +	struct repository *repo;
> +	struct oid_array current_batch;
> +	size_t batch_size;
> +};
> +
> +static void clear_backfill_context(struct backfill_context *ctx)
> +{
> +	oid_array_clear(&ctx->current_batch);
> +}

Nit: our style guide says that this should rather be
`backfill_context_clear()`.

> +static void download_batch(struct backfill_context *ctx)
> +{
> +	promisor_remote_get_direct(ctx->repo,
> +				   ctx->current_batch.oid,
> +				   ctx->current_batch.nr);
> +	oid_array_clear(&ctx->current_batch);
> +
> +	/*
> +	 * We likely have a new packfile. Add it to the packed list to
> +	 * avoid possible duplicate downloads of the same objects.
> +	 */
> +	reprepare_packed_git(ctx->repo);
> +}
> +
> +static int fill_missing_blobs(const char *path UNUSED,
> +			      struct oid_array *list,
> +			      enum object_type type,
> +			      void *data)
> +{
> +	struct backfill_context *ctx = data;
> +
> +	if (type != OBJ_BLOB)
> +		return 0;
> +
> +	for (size_t i = 0; i < list->nr; i++) {
> +		off_t size = 0;
> +		struct object_info info = OBJECT_INFO_INIT;
> +		info.disk_sizep = &size;
> +		if (oid_object_info_extended(ctx->repo,
> +					     &list->oid[i],
> +					     &info,
> +					     OBJECT_INFO_FOR_PREFETCH) ||
> +		    !size)
> +			oid_array_append(&ctx->current_batch, &list->oid[i]);
> +	}
> +
> +	if (ctx->current_batch.nr >= ctx->batch_size)
> +		download_batch(ctx);

Okay, so the batch size is just "best effort". If we walk a tree that
makes us exceed the batch size then we wouldn't issue a fetch during the
tree walk. Is there any specific reason for this behaviour?

In any case, as long as this is properly documented I think this should
be fine in general.

> +	return 0;
> +}
> +
> +static int do_backfill(struct backfill_context *ctx)
> +{
> +	struct rev_info revs;
> +	struct path_walk_info info = PATH_WALK_INFO_INIT;
> +	int ret;
> +
> +	repo_init_revisions(ctx->repo, &revs, "");
> +	handle_revision_arg("HEAD", &revs, 0, 0);
> +
> +	info.blobs = 1;
> +	info.tags = info.commits = info.trees = 0;
> +
> +	info.revs = &revs;
> +	info.path_fn = fill_missing_blobs;
> +	info.path_fn_data = ctx;
> +
> +	ret = walk_objects_by_path(&info);
> +
> +	/* Download the objects that did not fill a batch. */
> +	if (!ret)
> +		download_batch(ctx);
> +
> +	clear_backfill_context(ctx);

Are we leaking `revs` and `info`?

> +	return ret;
> +}
> +
>  int cmd_backfill(int argc, const char **argv, const char *prefix, struct repository *repo)
>  {
> +	struct backfill_context ctx = {
> +		.repo = repo,
> +		.current_batch = OID_ARRAY_INIT,
> +		.batch_size = 50000,
> +	};
>  	struct option options[] = {
>  		OPT_END(),
>  	};
> @@ -23,7 +123,5 @@ int cmd_backfill(int argc, const char **argv, const char *prefix, struct reposit
>  
>  	repo_config(repo, git_default_config, NULL);
>  
> -	die(_("not implemented"));
> -
> -	return 0;
> +	return do_backfill(&ctx);
>  }

The current iteration only backfills blobs as far as I can see. Do we
maybe want to keep the door open for future changes in git-backfill(1)
by implementing this via a "blob" subcommand?

Patrick

[gitgitgadget]

On the Git mailing list, Derrick Stolee wrote (reply to this):

On 12/16/24 3:01 AM, Patrick Steinhardt wrote:
> On Fri, Dec 06, 2024 at 08:07:15PM +0000, Derrick Stolee via GitGitGadget wrote:

>> +The `git backfill` command provides a way for the user to request that
>> +Git downloads the missing blobs (with optional filters) such that the
>> +missing blobs representing historical versions of files can be downloaded
>> +in batches. The `backfill` command attempts to optimize the request by
>> +grouping blobs that appear at the same path, hopefully leading to good
>> +delta compression in the packfile sent by the server.
>
> Hm. So we're asking the user to fix a usability issue of git-blame(1),
> don't we? Ideally, git-blame(1) itself should know to transparently
> batch the blobs it requires to compute its output, shouldn't it? That
> usecase alone doesn't yet convince me that git-backfill(1) is a good
> idea as I'd think we should rather fix the underlying issue.

I've looked into making this change for 'git blame' and it is a
nontrivial change. I'm not sure on the timeline that we could expect
'git blame' to be improved.

But you're right that this is not enough justification on its own. It's
an example of a kind of command that has these problems, including 'git
log [-p|-L]'.

> So are there other usecases for git-backfill(1)? I can imagine that it
> might be helpful in the context of scripts that know they'll operate on
> a bunch of blobs.

One major motivation is that it can essentially provide a way to do
resumable clones: get the commits and trees in one go with a blobless
clone, then download the blobs in batches. If something interrupts the
'git backfill' command, then restarting it will only repeat the most
recent batch.

Finally, in a later patch we can see that the --sparse option allows a
user to operate as if they have a full clone but where they only include
blob data within their sparse-checkout, providing reduced disk space and
network time to get to that state.

>> +	if (ctx->current_batch.nr >= ctx->batch_size)
>> +		download_batch(ctx);
>
> Okay, so the batch size is just "best effort". If we walk a tree that
> makes us exceed the batch size then we wouldn't issue a fetch during the
> tree walk. Is there any specific reason for this behaviour?
>
> In any case, as long as this is properly documented I think this should
> be fine in general.

The main reason is that we expect the server to return a packfile that
has many delta relationships within the objects at a given path. If we
split the batch in the middle of a path batch, then we are artificially
introducing breaks in the delta chains that could be wasteful.

However, this batching pattern could be problematic if there are a
million versions of a single file and the batch is too large to download
efficiently. This "absolute max batch size" is currently left as a
future extension.

>> +	clear_backfill_context(ctx);
>
> Are we leaking `revs` and `info`?

At the moment. Will fix.

>> +	return ret;
>> +}
>> +
>>   int cmd_backfill(int argc, const char **argv, const char *prefix, struct repository *repo)
>>   {
>> +	struct backfill_context ctx = {
>> +		.repo = repo,
>> +		.current_batch = OID_ARRAY_INIT,
>> +		.batch_size = 50000,
>> +	};
>>   	struct option options[] = {
>>   		OPT_END(),
>>   	};
>> @@ -23,7 +123,5 @@ int cmd_backfill(int argc, const char **argv, const char *prefix, struct reposit
>>
>>   	repo_config(repo, git_default_config, NULL);
>>
>> -	die(_("not implemented"));
>> -
>> -	return 0;
>> +	return do_backfill(&ctx);
>>   }
>
> The current iteration only backfills blobs as far as I can see. Do we
> maybe want to keep the door open for future changes in git-backfill(1)
> by implementing this via a "blob" subcommand?

I think that one tricky part is to ask "what could be missing?". With
Git's partial clone, it seems that we could have treeless or depth-
based tree restrictions. Technically, there could also be clones that
restrict to a set of sparse-checkout patterns, but I'm not aware of any
server that will respect those kinds of clones.  At the moment, the tree
walk would fault in any missing trees as they are seen, but this is
extremely inefficient.

I think that the path-walk API could be adjusted to be more careful to
check for the existence of an object before automatically loading it.
That would allow for batched downloads of missing trees, though a
second scan would be required to get the next layer of objects.

I'm not sure a subcommand is the right way to solve for this potential
future, but instead later we could adjust the logic to be better for
these treeless or tree-restricted clones.

Thanks,
-Stolee

[gitgitgadget]

On the Git mailing list, Patrick Steinhardt wrote (reply to this):

On Fri, Dec 06, 2024 at 08:07:17PM +0000, Derrick Stolee via GitGitGadget wrote:
> From: Derrick Stolee <[email protected]>
> 
> One way to significantly reduce the cost of a Git clone and later fetches is
> to use a blobless partial clone and combine that with a sparse-checkout that
> reduces the paths that need to be populated in the working directory. Not
> only does this reduce the cost of clones and fetches, the sparse-checkout
> reduces the number of objects needed to download from a promisor remote.
> 
> However, history investigations can be expensie as computing blob diffs will

s/expensie/expensive

Patrick

[gitgitgadget]

This patch series was integrated into seen via git@311b2d3.

[gitgitgadget]

This patch series was integrated into seen via git@b26b074.

[gitgitgadget]

This patch series was integrated into seen via git@aed51c8.