Re: [PATCH v9 0/9] Implement a batched fsync option for core.fsyncObjectFiles

[Neeraj Singh <nksingh85@gmail.com>]

On Tue, Nov 16, 2021 at 11:28 PM Ævar Arnfjörð Bjarmason
<avarab@gmail.com> wrote:
>
>
> On Tue, Nov 16 2021, Neeraj Singh wrote:
>
> > On Tue, Nov 16, 2021 at 12:10 AM Ævar Arnfjörð Bjarmason
> > <avarab@gmail.com> wrote:
> >>
> >>
> >> On Mon, Nov 15 2021, Neeraj K. Singh via GitGitGadget wrote:
> >>
> >> >  * Per [2], I'm leaving the fsyncObjectFiles configuration as is with
> >> >    'true', 'false', and 'batch'. This makes using old and new versions of
> >> >    git with 'batch' mode a little trickier, but hopefully people will
> >> >    generally be moving forward in versions.
> >> >
> >> > [1] See
> >> > https://lore.kernel.org/git/pull.1067.git.1635287730.gitgitgadget@gmail.com/
> >> > [2] https://lore.kernel.org/git/xmqqh7cimuxt.fsf@gitster.g/
> >>
> >> I really think leaving that in-place is just being unnecessarily
> >> cavalier. There's a lot of mixed-version environments where git is
> >> deployed in, and we almost never break the configuration in this way (I
> >> think in the past always by mistake).
> >
> >> In this case it's easy to avoid it, and coming up with a less narrow
> >> config model[1] seems like a good idea in any case to unify the various
> >> outstanding work in this area.
> >>
> >> More generally on this series, per the thread ending in [2] I really
> >
> > My primary goal in all of these changes is to move git-for-windows over to
> > a default of batch fsync so that it can get closer to other platforms
> > in performance
> > of 'git add' while still retaining the same level of data integrity.
> > I'm hoping that
> > most end-users are just sticking to defaults here.
> >
> > I'm happy to change the configuration schema again if there's a
> > consensus from the Git
> > community that backwards-compatibility of the configuration is
> > actually important to someone.
> >
> > Also, if we're doing a deeper rethink of the fsync configuration (as
> > prompted by this work and
> > Eric Wong's and Patrick Steinhardts work), do we want to retain a mode
> > where we fsync some
> > parts of the persistent repo data but not others?  If we add fsyncing
> > of the index in addition to the refs,
> > I believe we would have covered all of the critical data structures
> > that would be needed to find the
> > data that a user has added to the repo if they complete a series of
> > git commands and then experience
> > a system crash.
>
> Just talking about it is how we'll find consensus, maybe you & Junio
> would like to keep it as-is. I don't see why we'd expose this bad edge
> case in configuration handling to users when it's entirely avoidable,
> and we're still in the design phase.

After trying to figure out an implementation, I have a new proposal,
which I've shared on the other thread [1].

[1] https://lore.kernel.org/git/CANQDOdcdhfGtPg0PxpXQA5gQ4x9VknKDKCCi4HEB0Z1xgnjKzg@mail.gmail.com/

>
> >> don't get why we have code like this:
> >>
> >>         @@ -503,10 +504,12 @@ static void unpack_all(void)
> >>                 if (!quiet)
> >>                         progress = start_progress(_("Unpacking objects"), nr_objects);
> >>                 CALLOC_ARRAY(obj_list, nr_objects);
> >>         +       plug_bulk_checkin();
> >>                 for (i = 0; i < nr_objects; i++) {
> >>                         unpack_one(i);
> >>                         display_progress(progress, i + 1);
> >>                 }
> >>         +       unplug_bulk_checkin();
> >>                 stop_progress(&progress);
> >>
> >>                 if (delta_list)
> >>
> >> As opposed to doing an fsync on the last object we're
> >> processing. I.e. why do we need the step of intentionally making the
> >> objects unavailable in the tmp-objdir, and creating a "cookie" file to
> >> sync at the start/end, as opposed to fsyncing on the last file (which
> >> we're writing out anyway).
> >>
> >> 1. https://lore.kernel.org/git/211110.86r1bogg27.gmgdl@evledraar.gmail.com/
> >> 2. https://lore.kernel.org/git/20211111000349.GA703@neerajsi-x1.localdomain/
> >
> > It's important to not expose an object's final name until its contents
> > have been fsynced
> > to disk. We want to ensure that wherever we crash, we won't have a
> > loose object that
> > Git may later try to open where the filename doesn't match the content
> > hash. I believe it's
> > okay for a given OID to be missing, since a later command could
> > recreate it, but an object
> > with a wrong hash looks like it would persist until we do a git-fsck.
>
> Yes, we handle that rather badly, as I mentioned in some other threads,
> but not doing the fsync on the last object v.s. a "cookie" file right
> afterwards seems like a hail-mary at best, no?
>

I'm not quite grasping what you're saying here. Are you saying that
using a dummy
file instead of one of the actual objects is less likely to produce
the desired outcome
on actual filesystem implementations?

> > I thought about figuring out how to sync the last object rather than some random
> > "cookie" file, but it wasn't clear to me how I'd figure out which
> > object is actually last
> > from library code in a way that doesn't burden each command with
> > somehow figuring
> > out its last object and communicating that. The 'cookie' approach
> > seems to lead to a cleaner
> > interface for callers.
>
> The above quoted code is looping through nr_objects isn't it? Can't a
> "do fsync" be passed down to unpack_one() when we process the last loose
> object?

Are you proposing that we do something different for unpack_objects
versus update_index
and git-add?  I was hoping to keep all of the users of the batch fsync
functionality equivalent.
For the git-add workflow and update-index, we'd need to track the most
recent file so that we
can go back and fsync it.  I don't believe that syncing the last
object composes well with the existing
implementation of those commands.

Thanks,
Neeraj