Re: [PATCH v2] send-pack: never fetch when checking exclusions

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, Oct 11, 2019 at 03:08:22PM -0700, Jonathan Tan wrote:

> > As a general rule (and why I'm raising this issue in reply to Jonathan's
> > patch), I think most or all sites that want OBJECT_INFO_QUICK will want
> > SKIP_FETCH_OBJECT as well, and vice versa. The reasoning is generally
> > the same:
> > 
> >   - it's OK to racily have a false negative (we'll still be correct, but
> >     possibly a little less optimal)
> > 
> >   - it's expected and normal to be missing the object, so spending time
> >     double-checking the pack store wastes measurable time in real-world
> >     cases
> 
> I took a look on "next" and it's true for these reasons in most cases
> but not all.

Thanks for digging into this.

> QUICK implies SKIP_FETCH_OBJECT:
> 
>   fetch-pack.c: Run with fetch_if_missing=0 (from builtin/fetch.c,
>   builtin/fetch-pack.c, or through a lazy fetch) so OK.
>   
>   builtin/index-pack.c: Run with fetch_if_missing=0, so OK.
>   
>   builtin/fetch.c: Run with fetch_if_missing=0, so OK.
>   
>   object-store.h, sha1-file.c: Definition and implementation of this
>   flag.

Right, I think going in this direction is pretty simple. Having been
marked with QUICK, they hit both of my points from above. And if we want
to avoid re-scanning the pack directory because of cost, we _definitely_
want to avoid making an expensive network call.

> Everything is OK here. Now, SKIP_FETCH_OBJECT implies QUICK:
> 
>   cache-tree.c: I added this recently in f981ec18cf ("cache-tree: do not
>   lazy-fetch tentative tree", 2019-09-09). No problem with a false
>   negative, since we know how to reconstruct the tree. OK.
> [...]
>   send-pack.c: This patch (which is already in "next"). If we have a
>   false negative, we might accidentally send more than we need. But that
>   is not too bad.

Yeah, I think both of these could be QUICK.

>   promisor-remote.c: This is the slightly tricky one. We use this
>   information to determine if we got our lazily-fetched object from the
>   most recent lazy fetch, or if we should continue attempting to fetch the
>   given object from other promisor remotes; so this information is
>   important. However, adding QUICK doesn't lose us anything because the
>   lack of QUICK only helps us when there is another process packing
>   loose objects: if we got our object, our object will be in a pack
>   (because of the way the fetch is implemented - in particular, we need
>   a pack because we need the ".promisor" file).
> 
> So everything is OK except for promisor-remote.c, but even that is OK
> for another reason.

Yeah, though I wouldn't be sad to see that use a separate flag, since it
really is about promisor logic.

That implies to me maybe we should be using QUICK more aggressively, and
QUICK should auto-imply SKIP_FETCH_OBJECT.

> Having said that, perhaps we should consider promisor-remote.c as
> low-level code and expect it to know that objects are fetched into a
> packfile (as opposed to loose objects), so it can safely use QUICK
> (which is documented as checking packed after packed and loose). If no
> one disagrees, I can make such a patch after jt/push-avoid-lazy-fetch is
> merged to master (as is the plan, according to What's Cooking [1]).

I think it's OK to continue leaving out QUICK there if it's not causing
problems. It really is a bit different than the other cases.

-Peff



[Index of Archives]     [Linux Kernel Development]     [Gcc Help]     [IETF Annouce]     [DCCP]     [Netdev]     [Networking]     [Security]     [V4L]     [Bugtraq]     [Yosemite]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux SCSI]     [Fedora Users]

  Powered by Linux