git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH 2/3] index-pack: no blobs during outgoing link check

From
Patrick Steinhardt <ps@pks.im>
Date
Dec 3, 2024, 06:00 UTC
Message-ID
<Z06ejDgTnC6gWXgx@pks.im>
In-Reply-To
<300f53b8e39fa1dd55f65924d20f8abd22cbbfc9.1733170252.git.jonathantanmy@google.com>
On Mon, Dec 02, 2024 at 12:18:39PM -0800, Jonathan Tan wrote:
Show 12 quoted lines
> As a follow-up to the parent of this commit, it was found that not
> checking for the existence of blobs linked from trees sped up the fetch
> from 24m47.815s to 2m2.127s. Teach Git to do that.
> 
> The benefit of doing this is as above (fetch speedup), but the drawback
> is that if the packfile to be indexed references a local blob directly
> (that is, not through a local tree), that local blob is in danger of
> being garbage collected. Such a situation may arise if we push local
> commits, including one with a change to a blob in the root tree,
> and then the server incorporates them into its main branch through a
> "rebase" or "squash" merge strategy, and then we fetch the new main
> branch from the server.

Okay, so we know that we are basically doing the wrong thing with the optimization, but by skipping blobs we can get a significant speedup and the failure mode is that we will re-fetch the object in a later step. And because we think the situation is rare it shouldn't be a huge issue in practice.

Show 10 quoted lines
> This situation has not been observed yet - we have only noticed missing
> commits, not missing trees or blobs. (In fact, if it were believed that
> only missing commits are problematic, one could argue that we should
> also exclude trees during the outgoing link check; but it is safer to
> include them.)
> 
> Due to the rarity of the situation (it has not been observed to happen
> in real life), and because the "penalty" in such a situation is merely
> to refetch the missing blob when it's needed, the tradeoff seems
> worth it.

So is this a one-off event that may happen once per blob, or would we eventually evict the refetched blob and run into the same situation repeatedly?

Show 14 quoted lines
> diff --git a/builtin/index-pack.c b/builtin/index-pack.c
> index 8e7d14c17e..58d24540dc 100644
> --- a/builtin/index-pack.c
> +++ b/builtin/index-pack.c
> @@ -830,8 +830,10 @@ static void do_record_outgoing_links(struct object *obj)
>  			 * verified, so do not print any here.
>  			 */
>  			return;
> -		while (tree_entry_gently(&desc, &entry))
> -			record_outgoing_link(&entry.oid);
> +		while (tree_entry_gently(&desc, &entry)) {
> +			if (S_ISDIR(entry.mode))
> +				record_outgoing_link(&entry.oid);
> +		}

Without the context of the commit message this code snippet likely would not make any sense to a reader. The "correct" logic would be to record all objects, regardless of whether they are an object ID or not. But we explicitly choose not to as a tradeoff between performance and correctness.

All to say that we should have a comment here that explains what is going on.

Patrick
Previous: Jonathan TanNext: Jonathan Tan
Message 5 of 29 in “Performance improvements for repacking non-promisor objects”
  1. 0/3 Performance improvements for repacking non-promisor objectsJonathan Tan, Dec 2, 2024
  2. 1/3 index-pack: dedup first during outgoing link checkJonathan Tan, Dec 2, 2024
  3. Josh SteadmonDec 2, 2024
  4. 2/3 index-pack: no blobs during outgoing link checkJonathan Tan, Dec 2, 2024
  5. Patrick SteinhardtDec 3, 2024
  6. Jonathan TanDec 3, 2024
  7. Junio C HamanoDec 3, 2024
  8. 3/3 index-pack: commit tree during outgoing link checkJonathan Tan, Dec 2, 2024
  9. Junio C HamanoDec 3, 2024
  10. Jonathan TanDec 3, 2024
  11. Junio C HamanoDec 4, 2024
  12. Jonathan TanDec 9, 2024
  13. Junio C HamanoDec 9, 2024
  14. Josh SteadmonDec 2, 2024
  15. Junio C HamanoDec 3, 2024
  16. Junio C HamanoDec 3, 2024
  17. Junio C HamanoDec 3, 2024
  18. Junio C HamanoDec 3, 2024
  19. 0/3 Performance improvements for repacking non-promisor objectsJonathan Tan, Dec 3, 2024
  20. 1/3 index-pack --promisor: dedup before checking linksJonathan Tan, Dec 3, 2024
  21. 2/3 index-pack --promisor: don't check blobsJonathan Tan, Dec 3, 2024
  22. 3/3 index-pack --promisor: also check commits' treesJonathan Tan, Dec 3, 2024
  23. 0/3 Performance improvements for repacking non-promisor objectsJonathan Tan, Dec 3, 2024
  24. 1/3 index-pack --promisor: dedup before checking linksJonathan Tan, Dec 3, 2024
  25. Junio C HamanoDec 4, 2024
  26. 2/3 index-pack --promisor: don't check blobsJonathan Tan, Dec 3, 2024
  27. 3/3 index-pack --promisor: also check commits' treesJonathan Tan, Dec 3, 2024
  28. Junio C HamanoDec 4, 2024
  29. 4/3 index-pack: work around false positive use of uninitialized variableJunio C Hamano, Dec 4, 2024

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.