From: Jeff King <peff@peff.net>
To: Jonathan Tan <jonathantanmy@google.com>
Cc: steadmon@google.com, git@vger.kernel.org
Subject: Re: [PATCH] clone: do faster object check for partial clones
Date: Wed, 3 Apr 2019 21:33:44 -0400 [thread overview]
Message-ID: <20190404013343.GA4409@sigill.intra.peff.net> (raw)
In-Reply-To: <20190403205748.107979-1-jonathantanmy@google.com>
On Wed, Apr 03, 2019 at 01:57:48PM -0700, Jonathan Tan wrote:
> > This isn't strictly true, since we could get objects from elsewhere via
> > --shared or --reference. Those might not be promisor objects.
>
> I don't think local clones (which --shared or --reference implies) can
> be partial, but the bigger point is below.
Yeah, you're right about --shared. But I don't see any reason a
--reference clone could not be partial.
> > So it seems like this should be a feature of the child rev-list, to stop
> > walking the graph at any object that is in a promisor pack.
>
> We currently already do a less optimal version of this - we pass
> --exclude-promisor-objects to rev-list, which indeed stops traversal at
> any promisor objects (whether in a promisor pack or referenced by one).
> As far as I know, the problem is that to do so, we currently enumerate
> all the objects in all promisor packs, and all objects that those
> objects reference (which means we inflate them too) - so that we have an
> oidset that we can check objects against.
>
> A partial solution is for is_promisor_object() to first check if the
> given object is in a promisor pack, avoiding generating the set of
> promisor objects until necessary. This would work in a blob:none clone
> with the refs pointing all to commits or all to blobs, but would not
> work in a tree:none clone (or maybe, in this case, the clone would be
> small enough that performance is not a concern, hmm).
>
> Maybe the ideal solution is for rev-list to check if an object is in a
> promisor pack and if --exclude-promisor-objects is active, we do not
> follow any outgoing links.
I was thinking you could actually check it before even loading the
object. I.e., something like:
struct object_info oi = OBJECT_INFO_INIT;
if (!oid_object_info_extended(oid, &oi, 0) &&
oi->whence = OI_PACKED &&
oi->u.packed.pack->pack_promisor)) {
/*
* no point in even looking at its links,
* since the promisor pack claims that we
* can get anything we need later from the
* remote
*/
return 0; /* or whatever, depending where this goes ;) */
} else {
/* not a promisor object, load it and traverse as normal */
}
That doesn't quite work as an implementation of is_promisor_object(),
because it wouldn't know about items that we _don't_ have that are
promised. But I think it could work as part of the traversal in
list-objects.c, since we'd just be walking down a traversal from which
we presumably have all the objects.
I guess maybe it would be complicated if you had non-promisor objects
that refer indirectly to promisor ones. E.g., imagine ref A points to
object X, which is in a promisor pack pointing to Y (which we don't
have). But we also have ref B pointing to object Z which also refers to
Y, but _isn't_ in a promisor pack. I'm not sure that can actually happen
with the promisor mechanism, though (how did we get a Z without all of
its objects into a non-promisor pack?).
It's also a shame that it would incur an extra object lookup, since if
it _isn't_ in the promisor pack we'd then have to actually look it up
again via parse_object() or whatever. It may not be measurable though.
In an ideal world, we'd have an object access API that lets us open a
handle, ask things about it (like "which pack is this coming from") and
then load it if we want. But I don't think that needs to hold up this
particular topic.
-Peff
next prev parent reply other threads:[~2019-04-04 1:33 UTC|newest]
Thread overview: 27+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-04-03 17:27 [PATCH] clone: do faster object check for partial clones Josh Steadmon
2019-04-03 18:58 ` Jonathan Tan
2019-04-03 19:41 ` Jeff King
2019-04-03 20:57 ` Jonathan Tan
2019-04-04 0:21 ` Josh Steadmon
2019-04-04 1:33 ` Jeff King [this message]
2019-04-04 22:53 ` [PATCH v2] rev-list: exclude promisor objects at walk time Josh Steadmon
2019-04-04 23:08 ` Jeff King
2019-04-04 23:47 ` Josh Steadmon
2019-04-05 0:00 ` Jeff King
2019-04-05 0:09 ` Josh Steadmon
2019-04-08 20:59 ` Josh Steadmon
2019-04-08 21:06 ` [PATCH v3] " Josh Steadmon
2019-04-08 22:23 ` Christian Couder
2019-04-08 23:12 ` Josh Steadmon
2019-04-09 15:14 ` Junio C Hamano
2019-04-09 15:15 ` Jeff King
2019-04-09 15:43 ` Junio C Hamano
2019-04-09 16:35 ` Josh Steadmon
2019-04-09 18:04 ` SZEDER Gábor
2019-04-09 23:42 ` Josh Steadmon
2019-04-11 4:06 ` Jeff King
2019-04-12 22:38 ` Josh Steadmon
2019-04-13 5:34 ` Jeff King
2019-04-19 20:26 ` Josh Steadmon
2019-04-19 21:00 ` [PATCH v4] clone: do faster object check for partial clones Josh Steadmon
2019-04-22 21:31 ` Jeff King
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
List information: http://vger.kernel.org/majordomo-info.html
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20190404013343.GA4409@sigill.intra.peff.net \
--to=peff@peff.net \
--cc=git@vger.kernel.org \
--cc=jonathantanmy@google.com \
--cc=steadmon@google.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
Code repositories for project(s) associated with this public inbox
https://80x24.org/mirrors/git.git
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).