Regarding "git log" on "git series" metadata

Regarding "git log" on "git series" metadata

[Junio C Hamano]

After your talk at LPC2016, I was thinking about your proposal to
give an option to hide certain parents from "git log" traversal.

While I do not think we would terribly mind a new feature in the
core to support third-party additions like "git series" better, I
think this particular one is a big mistake that we shouldn't take.

For those listening from sidelines, here is a version of my
understanding of "git series":

 * "git series" wants to represent a patch series evolution.  It is
   a history of history, and each element of this evolution is
   represented by:

   - a commit object, that is used to describe what this reroll of
     the topic is about, and its parent links point at previous
     rerolls (it could be a merge of two independent incarnations of
     a series).

   - the tree contained in the commit object records the base commit
     where the topic forks from the main history, and the tip commit
     where the topic ends.  These are pointers into the main history
     DAG.

   - the tree may have other metadata, an example of which is the
     cover letter contents to be used when the topic becomes ready
     for re-submission.  There may be more metadata you would want
     to add in the future versions of "git series".

   Needless to say, the commits that represent the history of a
   series record a tree that is completely differently shaped.  The
   only relation between the series history and main history is that
   the former has pointers into the latter.

 * You chose to represent the base and tip commit object as gitlinks
   in the tree of a series commit, simply because it was a way that
   was already implemented to record a commit object name in a tree.

 * However, because gitlink is designed to be used for "external"
   things (the prominent example is submodule), recording these as
   gitlinks would guarantee that they will get GCed as a series
   progresses, the main history rewound and rewritten thereby making
   the base and tip recorded in the older part of the series history
   unreachable from the main history.  Because you want to make sure
   that base and tip objects will stay in the repository even after
   the topic branch in the main history gets rewound, this is not
   what you want.

 * In order to workaround that reachability issue, the hack you
   invented is to add the tip commit as a "parent" of a commit that
   represents one step in the series.  This may guarantee the
   reachability---as long as a commit in a series history is
   reachable from a ref, the tip and base commits will be reachable
   from there even if they are rebased away from the main history.
   But of course, there are downsides.

 * Due to this hack, feeding "gitk" (or "git log") a commit in the
   series history will give you nonsense results.  You are not
   interested in traversing or viewing the commits in the main
   history.

 * Because of the above, you propose another hack to tell the
   revision traversal machinery to optionally omit a parent commit
   that appear as a gitlink in the tree.

I think this is backwards.  The root cause of the issue you have
with "gitk" is because you added something that is *NOT* a parent to
your commit.  We shouldn't have to add a mechanism to filter
something that shouldn't have been added there in the first place.

I am wondering if an alternative approach would work better.

Imagine we invent a new tree entry type, "gitref", that is similar
to "gitlink" in that it can record a commit object name in a tree,
but unlike "gitlink" it does imply reachability.  And you do not add
phony parents to your commit object.  A tree that has "gitref"s in
it is about annotating the commits in the same repository (e.g. the
tree references two commits, "base" and "tip", to point into a slice
of the main history).  And it is perfectly sensible for such a
pointer to imply reachability---after all it serves different
purposes from "gitlink".

Another alternative that I am negative about (but is probably a
better hack than how you abused the "parent" link) might be to add a
new commit object header field that behaves similarly to "parent"
only in that it implies reachability.  But recording the extra
parent in commit object was not something you wanted to do in the
first place (i.e. your series processing is done solely on the
contents of the tree, and you do not read this extra parent). If you
need to add an in-tree reference to another commit in your future
versions of "git series", with either this variant or your original
implementation, you would end up needing adding more "parent" (or
pseudo parent) only to preserve reachability.  At that point, I
think it makes more sense to have entries in the tree to directly
ensure reachability, if you want these entries to always point at an
in-tree object.

I am afraid that I probably am two steps ahead of myself, because I
am reasonably sure that it is quite possible that I have overlooked
something trivially obvious that makes the "gitref" approach
unworkable.

Re: Regarding "git log" on "git series" metadata

[Jacob Keller]

On Fri, Nov 4, 2016 at 10:57 AM, Junio C Hamano <gitster@pobox.com> wrote:
> I think this is backwards.  The root cause of the issue you have
> with "gitk" is because you added something that is *NOT* a parent to
> your commit.  We shouldn't have to add a mechanism to filter
> something that shouldn't have been added there in the first place.
>
> I am wondering if an alternative approach would work better.
>
> Imagine we invent a new tree entry type, "gitref", that is similar
> to "gitlink" in that it can record a commit object name in a tree,
> but unlike "gitlink" it does imply reachability.  And you do not add
> phony parents to your commit object.  A tree that has "gitref"s in
> it is about annotating the commits in the same repository (e.g. the
> tree references two commits, "base" and "tip", to point into a slice
> of the main history).  And it is perfectly sensible for such a
> pointer to imply reachability---after all it serves different
> purposes from "gitlink".
>

I agree with your assessment here. The main difficulty in implementing
gitrefs is to ensure that they actually do get picked up by
reachability checks to prevent dropping commits. I'm not sure how easy
this is, but I would much rather we go this route rather than
continuing along with the hack. This seems like the ideal solution,
since it solves the entire problem and doesn't need more hacks bolted
on.

It would of course mean some work for people who previously used git
series as you would want to re-write the commits to drop the parent
links and become gitrefs instead of gitlinks. However, this can
(probably?) be solved by some sort of use of the filter-branch code.

I don't think you've hit upon any trivially obvious unworkable things.
It is probably somewhat complex to make the reachability checks detect
in-tree gitrefs but I don't think it would be impossible.

Thanks,
Jake

Re: Regarding "git log" on "git series" metadata

[Jeff King]

On Fri, Nov 04, 2016 at 12:19:55PM -0700, Jacob Keller wrote:

> I agree with your assessment here. The main difficulty in implementing
> gitrefs is to ensure that they actually do get picked up by
> reachability checks to prevent dropping commits. I'm not sure how easy
> this is, but I would much rather we go this route rather than
> continuing along with the hack. This seems like the ideal solution,
> since it solves the entire problem and doesn't need more hacks bolted
> on.

I think the main complication is that the reachability rules are used
during object transfer. So you'd probably want to introduce some
protocol extension to say "I understand gitrefs", so that when one side
says "I have sha1 X and its reachable objects", we know whether they are
including gitrefs there. And likewise receivers with
transfer.fsckObjects may complain about the new gitref tree mode
(fortunately a new object type shouldn't be needed).

You might also want fallback rules for storing gitrefs on "old" servers
(e.g., backfilling gitrefs you need if the server didn't them in the
initial fetch). But I guess storing any gitrefs on such a server is
inherently dangerous, because the server might prune them at any time.

So perhaps a related question is: how can gitrefs be designed such that
existing servers reject them (rather than accepting the push and then
later throwing away half the data). It would be easy to notice in the
client during a push that we are sending gitrefs to a server which does
not claim that capability. But it seems more robust if it is the server
who decides "I will not accept these bogus objects".

I haven't thought all that hard about this. That's just my initial
thoughts on what sound hard. Tweaking the reachability code doesn't seem
all that bad; we already know all of the spots that care about
S_ISGITLINK(). It may even be that some of those spots work out of the
box (because gitlinks are usually about telling the graph-walking code
that we _don't_ care about reachability; we do by default for trees and
blobs).

I'd be surprised if all such sites work out of the box, though. Even if
they see "ah, sha1 X is referenced by tree Y and isn't a gitlink, and
therefore should be reachable", they need to also note that "X" is a
commit and recursively walk its objects.

-Peff

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Fri, Nov 04, 2016 at 03:49:07PM -0400, Jeff King wrote:
> On Fri, Nov 04, 2016 at 12:19:55PM -0700, Jacob Keller wrote:
> 
> > I agree with your assessment here. The main difficulty in implementing
> > gitrefs is to ensure that they actually do get picked up by
> > reachability checks to prevent dropping commits. I'm not sure how easy
> > this is, but I would much rather we go this route rather than
> > continuing along with the hack. This seems like the ideal solution,
> > since it solves the entire problem and doesn't need more hacks bolted
> > on.
> 
> I think the main complication is that the reachability rules are used
> during object transfer. So you'd probably want to introduce some
> protocol extension to say "I understand gitrefs", so that when one side
> says "I have sha1 X and its reachable objects", we know whether they are
> including gitrefs there. And likewise receivers with
> transfer.fsckObjects may complain about the new gitref tree mode
> (fortunately a new object type shouldn't be needed).
> 
> You might also want fallback rules for storing gitrefs on "old" servers
> (e.g., backfilling gitrefs you need if the server didn't them in the
> initial fetch). But I guess storing any gitrefs on such a server is
> inherently dangerous, because the server might prune them at any time.
> 
> So perhaps a related question is: how can gitrefs be designed such that
> existing servers reject them (rather than accepting the push and then
> later throwing away half the data). It would be easy to notice in the
> client during a push that we are sending gitrefs to a server which does
> not claim that capability. But it seems more robust if it is the server
> who decides "I will not accept these bogus objects".

This seems like the critical problem, here.  The parent hack I used in
git-series might be a hack, but it transparently works with old servers
and clients.  So, for instance, I can push a git-series ref to github,
with no changes required on github's part.  If git added gitrefs, and I
started using them in git-series, then that'd eliminate parent hack and
allow many standard git tools to work naturally on git-series commits
and history, but it'd also mean that people couldn't push git-series
commits to any server until that server updates git.

That said, I'd *love* to have gitrefs available, for a wide variety of
applications, and I can see an argument for introducing them and waiting
a few years for them to become universally available, similar to the
process gitlinks went through.

But I'd also love to have a backward-compatible solution.

- Josh Triplett

Re: Regarding "git log" on "git series" metadata

[Jacob Keller]

On Fri, Nov 4, 2016 at 2:55 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> That said, I'd *love* to have gitrefs available, for a wide variety of
> applications, and I can see an argument for introducing them and waiting
> a few years for them to become universally available, similar to the
> process gitlinks went through.
>
> But I'd also love to have a backward-compatible solution.
>
> - Josh Triplett

I think that you won't really find a backwards compatible solution
other than something like automatically generating refs for each point
of history. I know that gerrit does something like this by storing
each version in "refs/changes/id/version" or something along those
lines. I think this might actually be cleaner than your parent links
hack, and could be used as a fallback for when gitrefs don't work,
though you'd have to code exactly how to tell what to push to a
repository when pushing a series?

Thanks,
Jake

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Fri, Nov 04, 2016 at 04:37:34PM -0700, Jacob Keller wrote:
> On Fri, Nov 4, 2016 at 2:55 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> > That said, I'd *love* to have gitrefs available, for a wide variety of
> > applications, and I can see an argument for introducing them and waiting
> > a few years for them to become universally available, similar to the
> > process gitlinks went through.
> >
> > But I'd also love to have a backward-compatible solution.
> >
> > - Josh Triplett
> 
> I think that you won't really find a backwards compatible solution
> other than something like automatically generating refs for each point
> of history. I know that gerrit does something like this by storing
> each version in "refs/changes/id/version" or something along those
> lines. I think this might actually be cleaner than your parent links
> hack, and could be used as a fallback for when gitrefs don't work,
> though you'd have to code exactly how to tell what to push to a
> repository when pushing a series?

I'm not sure what the advantage of that would be, and it would mean that
if you ever have one branch without pushing the other(s), you'd get
severe time-delated breakage due to pruning.  (And if you pushed the
series without the other ref(s), its history would look right but then
you couldn't access the underlying versions of the patch series.)

One of my design goals was to *not* need a special "git series push" or
"git series pull"; you should just be able to use git push and git pull,
and you can set up normal refspecs.

That said, I could fairly easily generate the existing format with
artificial parent refs for backward compatibility, and provide a way to
use the new gitref-based storage format if you know that all your
servers and clients can handle it.  I'm also open to other suggestions
for how to make such a transition while still working with every git
server and git client that exists today.

Re: Regarding "git log" on "git series" metadata

[Jacob Keller]

On Fri, Nov 4, 2016 at 12:49 PM, Jeff King <peff@peff.net> wrote:
> I think the main complication is that the reachability rules are used
> during object transfer. So you'd probably want to introduce some
> protocol extension to say "I understand gitrefs", so that when one side
> says "I have sha1 X and its reachable objects", we know whether they are
> including gitrefs there. And likewise receivers with
> transfer.fsckObjects may complain about the new gitref tree mode
> (fortunately a new object type shouldn't be needed).
>
> You might also want fallback rules for storing gitrefs on "old" servers
> (e.g., backfilling gitrefs you need if the server didn't them in the
> initial fetch). But I guess storing any gitrefs on such a server is
> inherently dangerous, because the server might prune them at any time.
>

Is it possible currently for a protocol extension to result in "oh the
server doesn't support this so I'm going to stop pushing"? This would
be a rather hard transition, but it would at least ensure that pushing
to a server which doesn't support gitrefs would fail rather than
silently accept objects and then discard them later? I think this is
the only real transition unless we can make a change that old servers
object to already.

> So perhaps a related question is: how can gitrefs be designed such that
> existing servers reject them (rather than accepting the push and then
> later throwing away half the data). It would be easy to notice in the
> client during a push that we are sending gitrefs to a server which does
> not claim that capability. But it seems more robust if it is the server
> who decides "I will not accept these bogus objects".
>
> I haven't thought all that hard about this. That's just my initial
> thoughts on what sound hard. Tweaking the reachability code doesn't seem
> all that bad; we already know all of the spots that care about
> S_ISGITLINK(). It may even be that some of those spots work out of the
> box (because gitlinks are usually about telling the graph-walking code
> that we _don't_ care about reachability; we do by default for trees and
> blobs).

Right. I'm assuming tree objects don't get checked for invalid mode
already? If they do, we could just change the mode to something
unsupported currently. But... that seems like it might not be the case
because it requires checking every tree object coming in?

I'm not familiar with what sort of checking already exists... Thoughts?

>
> I'd be surprised if all such sites work out of the box, though. Even if
> they see "ah, sha1 X is referenced by tree Y and isn't a gitlink, and
> therefore should be reachable", they need to also note that "X" is a
> commit and recursively walk its objects.
>

They won't all work out of the box, but it shouldn't be much work to
do this part.

> -Peff

Re: Regarding "git log" on "git series" metadata

[Jeff King]

On Fri, Nov 04, 2016 at 04:34:34PM -0700, Jacob Keller wrote:

> > You might also want fallback rules for storing gitrefs on "old" servers
> > (e.g., backfilling gitrefs you need if the server didn't them in the
> > initial fetch). But I guess storing any gitrefs on such a server is
> > inherently dangerous, because the server might prune them at any time.
> 
> Is it possible currently for a protocol extension to result in "oh the
> server doesn't support this so I'm going to stop pushing"?

Yes, it would be easy for the client to abort if the server fails to
advertise a particular extension.

What I would worry about more is that "somehow" an older client gets
hold of history with a gitref, and then pushes it. It would be nice if
even an old server said "nope, I don't understand this and I won't take
it" rather than propagating the data to a server that will throw it
away.

> Right. I'm assuming tree objects don't get checked for invalid mode
> already? If they do, we could just change the mode to something
> unsupported currently. But... that seems like it might not be the case
> because it requires checking every tree object coming in?
> 
> I'm not familiar with what sort of checking already exists... Thoughts?

If the server sets receive.fsckObjects, then fsck_tree() runs and will
reject any non-standard mode. That option is not the default, though
some big hosters set it (GitHub does, but I am actually not that worried
about GitHub; if gitrefs support materialized I would probably ship it
there fairly promptly).

-Peff

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On November 4, 2016 7:48:17 PM MDT, Jeff King <peff@peff.net> wrote:
>On Fri, Nov 04, 2016 at 04:34:34PM -0700, Jacob Keller wrote:
>
>> > You might also want fallback rules for storing gitrefs on "old"
>servers
>> > (e.g., backfilling gitrefs you need if the server didn't them in
>the
>> > initial fetch). But I guess storing any gitrefs on such a server is
>> > inherently dangerous, because the server might prune them at any
>time.
>> 
>> Is it possible currently for a protocol extension to result in "oh
>the
>> server doesn't support this so I'm going to stop pushing"?
>
>Yes, it would be easy for the client to abort if the server fails to
>advertise a particular extension.

And the reverse (old client, new server) should work as well? 

- Josh Triplett

Re: Regarding "git log" on "git series" metadata

[Jeff King]

On Fri, Nov 04, 2016 at 09:55:23PM -0600, Josh Triplett wrote:

> >> Is it possible currently for a protocol extension to result in "oh
> >the
> >> server doesn't support this so I'm going to stop pushing"?
> >
> >Yes, it would be easy for the client to abort if the server fails to
> >advertise a particular extension.
> 
> And the reverse (old client, new server) should work as well?

Yes. The server says "I know about gitrefs" and if the client does not
respond with "I also know about gitrefs", then the server can act
appropriately (e.g., for a fetch, bail if the fetched content includes
gitrefs).

-Peff

Re: Regarding "git log" on "git series" metadata

[Junio C Hamano]

Jeff King <peff@peff.net> writes:

> On Fri, Nov 04, 2016 at 12:19:55PM -0700, Jacob Keller wrote:
>
>> I agree with your assessment here. The main difficulty in implementing
>> gitrefs is to ensure that they actually do get picked up by
>> reachability checks to prevent dropping commits. I'm not sure how easy
>> this is, but I would much rather we go this route rather than
>> continuing along with the hack. This seems like the ideal solution,
>> since it solves the entire problem and doesn't need more hacks bolted
>> on.
>
> I think the main complication is that the reachability rules are used
> during object transfer. So you'd probably want to introduce some
> protocol extension to say "I understand gitrefs", so that when one side
> says "I have sha1 X and its reachable objects", we know whether they are
> including gitrefs there. And likewise receivers with
> transfer.fsckObjects may complain about the new gitref tree mode
> (fortunately a new object type shouldn't be needed).

Quite honestly I do not think backward compatibility here matters.
When gitlinks were introduced, a repository that was created with
gitlink capable version of Git would have failed "git fsck" that is
not gitlink aware, and I think this new "link with reachability" is
the same deal.  No existing implemention understands a tree entry
whose mode bits are 140000 or whatever new bit pattern we would
assign to this thing.  You have to wait until both ends understand
the new thing, and that is perfectly OK.

Besides, I think the point of having this discussion is that Josh
did a good prototyping work of "git series" to discover what he can
do in that area of "keeping track of history of history" and what
operations are useful, without wasting time on mucking with the
object model and traversal machinery that are available to him.  

Now his prototyping is at the point where he knows at least one
enhancement to the core that would help him to redo the prototype in
the right way.  And I do not mind helping him from the core side.

Re: Regarding "git log" on "git series" metadata

[Jeff King]

On Fri, Nov 04, 2016 at 09:41:06PM -0700, Junio C Hamano wrote:

> > I think the main complication is that the reachability rules are used
> > during object transfer. So you'd probably want to introduce some
> > protocol extension to say "I understand gitrefs", so that when one side
> > says "I have sha1 X and its reachable objects", we know whether they are
> > including gitrefs there. And likewise receivers with
> > transfer.fsckObjects may complain about the new gitref tree mode
> > (fortunately a new object type shouldn't be needed).
> 
> Quite honestly I do not think backward compatibility here matters.
> When gitlinks were introduced, a repository that was created with
> gitlink capable version of Git would have failed "git fsck" that is
> not gitlink aware, and I think this new "link with reachability" is
> the same deal.  No existing implemention understands a tree entry
> whose mode bits are 140000 or whatever new bit pattern we would
> assign to this thing.  You have to wait until both ends understand
> the new thing, and that is perfectly OK.

I'm OK with saying "if you use the gitref feature, you cannot push or
pull those objects with remotes that do not understand it".  But unlike
gitlink, if we fail to notice the situation, we run into a case where we
might silently lose objects, which is bad. So I think we need to be a
bit more careful.

I don't think the problems are insurmountable. I just think that's where
the real complexity is, not in the changes to teach a single git about
gitrefs.

I'm happy to stand back and let you or Josh figure out all the corner
cases. :)

-Peff

Re: Regarding "git log" on "git series" metadata

[Junio C Hamano]

Junio C Hamano <gitster@pobox.com> writes:

>> I think the main complication is that the reachability rules are used
>> during object transfer.

One should not type after spending 20+ waking hours on plane and
airport.  I missed it when I wrote my first response, but yes, the
reachability that originates from inside a tree object indeed is a
problem, as we do not dig into trees while doing the want/have/ack
exhange.

Re: Regarding "git log" on "git series" metadata

[Christian Couder]

On Fri, Nov 4, 2016 at 6:57 PM, Junio C Hamano <gitster@pobox.com> wrote:
>
> Imagine we invent a new tree entry type, "gitref", that is similar
> to "gitlink" in that it can record a commit object name in a tree,
> but unlike "gitlink" it does imply reachability.  And you do not add
> phony parents to your commit object.  A tree that has "gitref"s in
> it is about annotating the commits in the same repository (e.g. the
> tree references two commits, "base" and "tip", to point into a slice
> of the main history).  And it is perfectly sensible for such a
> pointer to imply reachability---after all it serves different
> purposes from "gitlink".

The more I think about this (and also about how to limit ref
advertisements as recently discussed in
https://public-inbox.org/git/20161024132932.i42rqn2vlpocqmkq@sigill.intra.peff.net/),
the more I think about Shawn's RefTree:

https://public-inbox.org/git/CAJo=hJvnAPNAdDcAAwAvU9C4RVeQdoS3Ev9WTguHx4fD0V_nOg@mail.gmail.com/

Couldn't a RefTree be used to store refs that point to the base
commit, the tip commit and the blob that contains the cover letter,
and maybe also a ref pointing to the RefTree of the previous version
of the series?

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Fri, Nov 04, 2016 at 09:47:41PM +0100, Christian Couder wrote:
> On Fri, Nov 4, 2016 at 6:57 PM, Junio C Hamano <gitster@pobox.com> wrote:
> >
> > Imagine we invent a new tree entry type, "gitref", that is similar
> > to "gitlink" in that it can record a commit object name in a tree,
> > but unlike "gitlink" it does imply reachability.  And you do not add
> > phony parents to your commit object.  A tree that has "gitref"s in
> > it is about annotating the commits in the same repository (e.g. the
> > tree references two commits, "base" and "tip", to point into a slice
> > of the main history).  And it is perfectly sensible for such a
> > pointer to imply reachability---after all it serves different
> > purposes from "gitlink".
> 
> The more I think about this (and also about how to limit ref
> advertisements as recently discussed in
> https://public-inbox.org/git/20161024132932.i42rqn2vlpocqmkq@sigill.intra.peff.net/),
> the more I think about Shawn's RefTree:
> 
> https://public-inbox.org/git/CAJo=hJvnAPNAdDcAAwAvU9C4RVeQdoS3Ev9WTguHx4fD0V_nOg@mail.gmail.com/
> 
> Couldn't a RefTree be used to store refs that point to the base
> commit, the tip commit and the blob that contains the cover letter,
> and maybe also a ref pointing to the RefTree of the previous version
> of the series?

That's really interesting!  The Software Heritage project is working on
something similar, because they want to store all the refs as part of
their data model as well.  I'll point them to the reftree work.

If upstream git supported RefTree, I could potentially use that for
git-series.  However, I do want a commit message and history for the
series itself, and using refs in the reftree to refer to the parents
seems like abusing reftree to recreate commits, in a reversal of the
hack of using commit parents as a reftree. :)

What if, rather than storing a hash reference to a reftree as a single
reference and replacing it with no history, a reftree could be
referenced from a commit and have history?  (That would also allow
tagging a version of the reftree.)

Re: Regarding "git log" on "git series" metadata

[Christian Couder]

On Fri, Nov 4, 2016 at 10:19 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> On Fri, Nov 04, 2016 at 09:47:41PM +0100, Christian Couder wrote:
>> On Fri, Nov 4, 2016 at 6:57 PM, Junio C Hamano <gitster@pobox.com> wrote:
>> >
>> > Imagine we invent a new tree entry type, "gitref", that is similar
>> > to "gitlink" in that it can record a commit object name in a tree,
>> > but unlike "gitlink" it does imply reachability.  And you do not add
>> > phony parents to your commit object.  A tree that has "gitref"s in
>> > it is about annotating the commits in the same repository (e.g. the
>> > tree references two commits, "base" and "tip", to point into a slice
>> > of the main history).  And it is perfectly sensible for such a
>> > pointer to imply reachability---after all it serves different
>> > purposes from "gitlink".
>>
>> The more I think about this (and also about how to limit ref
>> advertisements as recently discussed in
>> https://public-inbox.org/git/20161024132932.i42rqn2vlpocqmkq@sigill.intra.peff.net/),
>> the more I think about Shawn's RefTree:
>>
>> https://public-inbox.org/git/CAJo=hJvnAPNAdDcAAwAvU9C4RVeQdoS3Ev9WTguHx4fD0V_nOg@mail.gmail.com/
>>
>> Couldn't a RefTree be used to store refs that point to the base
>> commit, the tip commit and the blob that contains the cover letter,
>> and maybe also a ref pointing to the RefTree of the previous version
>> of the series?
>
> That's really interesting!  The Software Heritage project is working on
> something similar, because they want to store all the refs as part of
> their data model as well.  I'll point them to the reftree work.

Yeah, I know them :-) and I think I have already told Stefano
Zacchiroli about this, but I am not sure anymore.
Anyway I am CC'ing him.

> If upstream git supported RefTree, I could potentially use that for
> git-series.  However, I do want a commit message and history for the
> series itself, and using refs in the reftree to refer to the parents
> seems like abusing reftree to recreate commits, in a reversal of the
> hack of using commit parents as a reftree. :)

Yeah, maybe :-) But the properties of the existing Git objects we
already use wouldn't change at all.

> What if, rather than storing a hash reference to a reftree as a single
> reference and replacing it with no history,

In what I suggest the history is kept because the new reftree has a
ref that points to the old one it is replacing.

Yeah, this reftree history maybe seen as "redundant" with the commit
history, but in my opinion this can be seen as a "feature" that will
prevent us from "mucking" too much with the commit object.

> a reftree could be
> referenced from a commit and have history?  (That would also allow
> tagging a version of the reftree.)

I think that tags are already allowed to point to any kind of Git
object, so tagging a reftree should be allowed anyway if we add a
reftree object.

Re: Regarding "git log" on "git series" metadata

[Stefano Zacchiroli]

Hi everyone,

On Sat, Nov 05, 2016 at 12:04:08AM +0100, Christian Couder wrote:
> On Fri, Nov 4, 2016 at 10:19 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> > On Fri, Nov 04, 2016 at 09:47:41PM +0100, Christian Couder wrote:
> >>
> >> Couldn't a RefTree be used to store refs that point to the base
> >> commit, the tip commit and the blob that contains the cover letter,
> >> and maybe also a ref pointing to the RefTree of the previous version
> >> of the series?
> >
> > That's really interesting!  The Software Heritage project is working on
> > something similar, because they want to store all the refs as part of
> > their data model as well.  I'll point them to the reftree work.
> 
> Yeah, I know them :-) and I think I have already told Stefano
> Zacchiroli about this, but I am not sure anymore.
> Anyway I am CC'ing him.

Thanks Christian (and Josh, on swh-devel) for pointing me to this.

As a bit of background, the conceptual data model we have adopted for
Software Heritage [1] is indeed that of a global Merkle DAG, very much
inspired by Git, but where we deduplicate past the boundaries of
individual VCS repositories. This way we can store only once the same
software artifacts (blobs, trees, commits, etc.) even when they can be
found at different software origins [2] (be it due to GitHub-like forks,
projects moving around, or simply rogue copies of the same code
scattered around the Internet).

[1]: https://www.softwareheritage.org/

[2]: "software origin" is Software Heritage terminology, which just
     stands for places on the Internet where we can find source code

In our original design the topmost entries in our Merkle hierarchy used
to be commits and tags, similar to what Git does. But then we realized
that doing so inhibited us from sharing entire repository states across
multiple software origins or multiple visits of the same software
origin.  So we decided to add "repository snapshot objects" as our
topmost entries, which are essentially git-like objects that map refs to
the ID of the corresponding (typed-)objects. Rationale and a more
lengthy description of this is available on our wiki [3]. It is not
implemented yet, but we're pretty sold on the design at this point.

[3]: https://wiki.softwareheritage.org/index.php?title=Repository_snapshot_objects

Now, even if my only awareness of what's going on in Git upstream is
limited to sporadic chats with Josh and Christian :-), it seems to me
that various ideas in the Git ecosystem go in the same direction of our
snapshot objects (git-series, RefTree). Which is understandable, given a
number of use cases might be served by this.

I don't think we have much to contribute to discussion or implementation
here, and for our needs it doesn't really matter which one gets
implemented. That's because we need an implementation of the concept
which is *external* to Git anyhow. But even if it happens to exist
within actual VCS, it's not a big deal for us, as we do have ways to
distinguish "synthetic" objects in the DAG that we create for our own
needs from "real" objects coming from actual software origins.
(Another example of this concept we already have is when we inject
distribution source packages or tarballs in our archive. In that case we
create synthetic commits that points to the tree extracted from the
tarball/package, preserving the ability to distinguish them from real
commits coming from VCS out there.)

If you think we can help in any other way, other than sharing our
experiences and design considerations that is, please let me know! (I'm
not subscribed to the Git upstream mailing list, but feel free to Cc:-me
in conversations related to this topic.)

Cheers.
-- 
Stefano Zacchiroli . zack@upsilon.cc . upsilon.cc/zack . . o . . . o . o
Computer Science Professor . CTO Software Heritage . . . . . o . . . o o
Former Debian Project Leader . OSI Board Director  . . . o o o . . . o .
« the first rule of tautology club is the first rule of tautology club »

Re: Regarding "git log" on "git series" metadata

[Christian Couder]

On Fri, Nov 4, 2016 at 10:19 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> On Fri, Nov 04, 2016 at 09:47:41PM +0100, Christian Couder wrote:
>> On Fri, Nov 4, 2016 at 6:57 PM, Junio C Hamano <gitster@pobox.com> wrote:
>> >
>> > Imagine we invent a new tree entry type, "gitref", that is similar
>> > to "gitlink" in that it can record a commit object name in a tree,
>> > but unlike "gitlink" it does imply reachability.  And you do not add
>> > phony parents to your commit object.  A tree that has "gitref"s in
>> > it is about annotating the commits in the same repository (e.g. the
>> > tree references two commits, "base" and "tip", to point into a slice
>> > of the main history).  And it is perfectly sensible for such a
>> > pointer to imply reachability---after all it serves different
>> > purposes from "gitlink".
>>
>> The more I think about this (and also about how to limit ref
>> advertisements as recently discussed in
>> https://public-inbox.org/git/20161024132932.i42rqn2vlpocqmkq@sigill.intra.peff.net/),
>> the more I think about Shawn's RefTree:
>>
>> https://public-inbox.org/git/CAJo=hJvnAPNAdDcAAwAvU9C4RVeQdoS3Ev9WTguHx4fD0V_nOg@mail.gmail.com/

Just to make things clear, after reading the above link that I posted :-) ...

>> Couldn't a RefTree be used to store refs that point to the base
>> commit, the tip commit and the blob that contains the cover letter,
>> and maybe also a ref pointing to the RefTree of the previous version
>> of the series?
>
> That's really interesting!  The Software Heritage project is working on
> something similar, because they want to store all the refs as part of
> their data model as well.  I'll point them to the reftree work.
>
> If upstream git supported RefTree, I could potentially use that for
> git-series.  However, I do want a commit message and history for the
> series itself, and using refs in the reftree to refer to the parents
> seems like abusing reftree to recreate commits, in a reversal of the
> hack of using commit parents as a reftree. :)
>
> What if, rather than storing a hash reference to a reftree as a single
> reference and replacing it with no history, a reftree could be
> referenced from a commit and have history?  (That would also allow
> tagging a version of the reftree.)

... I think that indeed that's what Shawn's reftree proposal is about,
so I agree that it makes sense.

We just need to find a good way to specify object reachability.

Re: Regarding "git log" on "git series" metadata

[Junio C Hamano]

Christian Couder <christian.couder@gmail.com> writes:

> Couldn't a RefTree be used to store refs that point to the base
> commit,

I think it is the other way around.  With the new "gitref" thing
that is a pointer to an in-repository commit, RefTree can be
naturally implemented.

Re: Regarding "git log" on "git series" metadata

[Christian Couder]

On Sat, Nov 5, 2016 at 5:42 AM, Junio C Hamano <gitster@pobox.com> wrote:
> Christian Couder <christian.couder@gmail.com> writes:
>
>> Couldn't a RefTree be used to store refs that point to the base
>> commit,
>
> I think it is the other way around.  With the new "gitref" thing
> that is a pointer to an in-repository commit, RefTree can be
> naturally implemented.

Yeah, I should have read Shawn's RefTree email thread again before
posting and especially before replying to Josh.

Re: Regarding "git log" on "git series" metadata

[Christian Couder]

On Sat, Nov 5, 2016 at 1:17 PM, Christian Couder
<christian.couder@gmail.com> wrote:
> On Sat, Nov 5, 2016 at 5:42 AM, Junio C Hamano <gitster@pobox.com> wrote:
>> Christian Couder <christian.couder@gmail.com> writes:
>>
>>> Couldn't a RefTree be used to store refs that point to the base
>>> commit,
>>
>> I think it is the other way around.  With the new "gitref" thing
>> that is a pointer to an in-repository commit, RefTree can be
>> naturally implemented.
>
> Yeah, I should have read Shawn's RefTree email thread again before
> posting and especially before replying to Josh.

By the way, reading the following email by Peff where gitlink
reachability was already discussed:

https://public-inbox.org/git/20151217221045.GA8150@sigill.intra.peff.net/

and where Peff wrote:

> Of course, the lack of reachability has advantages, too. You can
> drop commits pointed to by old reflogs without rewriting the ref
> history. Unfortunately you cannot expunge the reflogs at all. That's
> good if you like audit trails. Bad if you are worried that your reflogs
> will grow large. :)

I think that we may not need "gitref" at all. We perhaps could just
have more ways to configure and tweak how a repo manages commit
reachability related to gitlinks.

With shallow clones we already need ways to configure and tweak commit
reachability anyway.

And with what Peff says above it looks like we will need ways
configure and tweak commit reachability with gitlink/gitref anyway. So
the point of gitref compared to gitlink would be that they just have a
different reachability by default. But couldn't that be replaced by a
default rule saying that when a gitlink is reached "this way or that
way" then the commit reachability should be enforced, and otherwise it
should not be?

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Sat, Nov 05, 2016 at 01:45:27PM +0100, Christian Couder wrote:
> And with what Peff says above it looks like we will need ways
> configure and tweak commit reachability with gitlink/gitref anyway. So
> the point of gitref compared to gitlink would be that they just have a
> different reachability by default. But couldn't that be replaced by a
> default rule saying that when a gitlink is reached "this way or that
> way" then the commit reachability should be enforced, and otherwise it
> should not be?

Any version of git unaware of that rule, though, would consider objects
only reachable by gitlink as unreachable and delete them, causing data
loss.  Likewise for a server not aware of that rule.  And a server
unaware of that rule would not supply those objects to a client pulling
such a branch.

So I don't think "gitlink defined as reachable" quite works, unless we
make some other format-incompatible change that forces clients and
servers touching that gitlink to know about that reachability rule.  (In
the absence of a hack such as making the same commit a parent.)

Re: Regarding "git log" on "git series" metadata

[Christian Couder]

On Sat, Nov 5, 2016 at 4:18 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> On Sat, Nov 05, 2016 at 01:45:27PM +0100, Christian Couder wrote:
>> And with what Peff says above it looks like we will need ways
>> configure and tweak commit reachability with gitlink/gitref anyway. So
>> the point of gitref compared to gitlink would be that they just have a
>> different reachability by default. But couldn't that be replaced by a
>> default rule saying that when a gitlink is reached "this way or that
>> way" then the commit reachability should be enforced, and otherwise it
>> should not be?
>
> Any version of git unaware of that rule, though, would consider objects
> only reachable by gitlink as unreachable and delete them, causing data
> loss.  Likewise for a server not aware of that rule.  And a server
> unaware of that rule would not supply those objects to a client pulling
> such a branch.

Yeah, so you would really need an up-to-date server and client to
store the git-series data.
But anyway if we create a gitref object, you would also need
up-to-date servers and clients to make it work.

> So I don't think "gitlink defined as reachable" quite works, unless we
> make some other format-incompatible change that forces clients and
> servers touching that gitlink to know about that reachability rule.  (In
> the absence of a hack such as making the same commit a parent.)

There are other tools that would like to tweak reachability rules for objects.

For example Mike Hommey's git-cinnabar:

  https://public-inbox.org/git/20150331100756.GA13377@glandium.org/

and my current work on external object databases:

  https://public-inbox.org/git/20160628181933.24620-1-chriscool@tuxfamily.org/

would be interested in a way to make some blobs not reachable.

So if we had default rules and a generic way to specify that some
objects are, or are not, reachable, that could be used by many tools,
and the design of these tools would be simplified.

Maybe this could be specified in the attributes files, or in a special
file like for shallow clone.

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Sat, Nov 05, 2016 at 09:21:58PM +0100, Christian Couder wrote:
> On Sat, Nov 5, 2016 at 4:18 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> > On Sat, Nov 05, 2016 at 01:45:27PM +0100, Christian Couder wrote:
> >> And with what Peff says above it looks like we will need ways
> >> configure and tweak commit reachability with gitlink/gitref anyway. So
> >> the point of gitref compared to gitlink would be that they just have a
> >> different reachability by default. But couldn't that be replaced by a
> >> default rule saying that when a gitlink is reached "this way or that
> >> way" then the commit reachability should be enforced, and otherwise it
> >> should not be?
> >
> > Any version of git unaware of that rule, though, would consider objects
> > only reachable by gitlink as unreachable and delete them, causing data
> > loss.  Likewise for a server not aware of that rule.  And a server
> > unaware of that rule would not supply those objects to a client pulling
> > such a branch.
> 
> Yeah, so you would really need an up-to-date server and client to
> store the git-series data.
> But anyway if we create a gitref object, you would also need
> up-to-date servers and clients to make it work.

Agreed, but gitrefs have the advantage of failing safe, rather than
failing with dataloss.

- Josh Triplett

Re: Regarding "git log" on "git series" metadata

[Jacob Keller]

On Sat, Nov 5, 2016 at 1:25 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> On Sat, Nov 05, 2016 at 09:21:58PM +0100, Christian Couder wrote:
>> On Sat, Nov 5, 2016 at 4:18 PM, Josh Triplett <josh@joshtriplett.org> wrote:
>> > On Sat, Nov 05, 2016 at 01:45:27PM +0100, Christian Couder wrote:
>> >> And with what Peff says above it looks like we will need ways
>> >> configure and tweak commit reachability with gitlink/gitref anyway. So
>> >> the point of gitref compared to gitlink would be that they just have a
>> >> different reachability by default. But couldn't that be replaced by a
>> >> default rule saying that when a gitlink is reached "this way or that
>> >> way" then the commit reachability should be enforced, and otherwise it
>> >> should not be?
>> >
>> > Any version of git unaware of that rule, though, would consider objects
>> > only reachable by gitlink as unreachable and delete them, causing data
>> > loss.  Likewise for a server not aware of that rule.  And a server
>> > unaware of that rule would not supply those objects to a client pulling
>> > such a branch.
>>
>> Yeah, so you would really need an up-to-date server and client to
>> store the git-series data.
>> But anyway if we create a gitref object, you would also need
>> up-to-date servers and clients to make it work.
>
> Agreed, but gitrefs have the advantage of failing safe, rather than
> failing with dataloss.
>
> - Josh Triplett

Isn't the "failing safe" only true if the client disconnects when a
server doesn't advertise "i understand gitrefs"? So couldn't we, as
part of the rules for reachability advertise a capability that does a
similar thing and fails safe as well?

Thanks.
Jake

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Sat, Nov 05, 2016 at 09:50:07PM -0700, Jacob Keller wrote:
> On Sat, Nov 5, 2016 at 1:25 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> > On Sat, Nov 05, 2016 at 09:21:58PM +0100, Christian Couder wrote:
> >> On Sat, Nov 5, 2016 at 4:18 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> >> > On Sat, Nov 05, 2016 at 01:45:27PM +0100, Christian Couder wrote:
> >> >> And with what Peff says above it looks like we will need ways
> >> >> configure and tweak commit reachability with gitlink/gitref anyway. So
> >> >> the point of gitref compared to gitlink would be that they just have a
> >> >> different reachability by default. But couldn't that be replaced by a
> >> >> default rule saying that when a gitlink is reached "this way or that
> >> >> way" then the commit reachability should be enforced, and otherwise it
> >> >> should not be?
> >> >
> >> > Any version of git unaware of that rule, though, would consider objects
> >> > only reachable by gitlink as unreachable and delete them, causing data
> >> > loss.  Likewise for a server not aware of that rule.  And a server
> >> > unaware of that rule would not supply those objects to a client pulling
> >> > such a branch.
> >>
> >> Yeah, so you would really need an up-to-date server and client to
> >> store the git-series data.
> >> But anyway if we create a gitref object, you would also need
> >> up-to-date servers and clients to make it work.
> >
> > Agreed, but gitrefs have the advantage of failing safe, rather than
> > failing with dataloss.
> >
> > - Josh Triplett
> 
> Isn't the "failing safe" only true if the client disconnects when a
> server doesn't advertise "i understand gitrefs"? So couldn't we, as
> part of the rules for reachability advertise a capability that does a
> similar thing and fails safe as well?

We could, but if we (or one of the many third-party git implementations)
miss a case, gitlinks+reachability may appear to work in many cases with
dataloss afterward, while gitrefs will fail early and not appear
functional.

Re: Regarding "git log" on "git series" metadata

[Junio C Hamano]

Josh Triplett <josh@joshtriplett.org> writes:

> We could, but if we (or one of the many third-party git implementations)
> miss a case, gitlinks+reachability may appear to work in many cases with
> dataloss afterward, while gitrefs will fail early and not appear
> functional.

I wonder what happens if we do not introduce the "gitref" but
instead change the behaviour of "gitlink" to imply an optional
reachability.  That is, when enumerating what is reachable in your
repository, if you see a gitlink and if you notice that you locally
have the target of that gitlink, you follow, but if you know you
lack it, you do not error out.  This may be making things too
complex to feasibily implement by simplify them ;-) and I see a few
immediate fallout that needs to be thought through (i.e. downsides)
and a few upsides, too.  I am feeling feverish and not thinking
straight, so I won't try to weigh pros-and-cons.  

This would definitely need protocol extension when transferring
objects across repositories.

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Sun, Nov 06, 2016 at 09:14:56AM -0800, Junio C Hamano wrote:
> Josh Triplett <josh@joshtriplett.org> writes:
> > We could, but if we (or one of the many third-party git implementations)
> > miss a case, gitlinks+reachability may appear to work in many cases with
> > dataloss afterward, while gitrefs will fail early and not appear
> > functional.
> 
> I wonder what happens if we do not introduce the "gitref" but
> instead change the behaviour of "gitlink" to imply an optional
> reachability.  That is, when enumerating what is reachable in your
> repository, if you see a gitlink and if you notice that you locally
> have the target of that gitlink, you follow, but if you know you
> lack it, you do not error out.  This may be making things too
> complex to feasibily implement by simplify them ;-) and I see a few
> immediate fallout that needs to be thought through (i.e. downsides)
> and a few upsides, too.  I am feeling feverish and not thinking
> straight, so I won't try to weigh pros-and-cons.  
> 
> This would definitely need protocol extension when transferring
> objects across repositories.

It'd also need a repository format extension locally.  Otherwise, if you
ever touched that repository with an older git (or a tool built on an
older libgit2 or JGit or other library), you could lose data.

It does seem conceptually appealing, though.  In an ideal world, the
original version of gitlink would have had opt-out reachability (and
.gitmodules with an external repository reference could count as opting
out).

But I can't think of any case where it's OK for a git implementation to
not know about this reachability extension and still operate on the
gitlink.  And given that, it might as well use a new object type that
the old version definitely won't think it understands.

- Josh Triplett

Re: Regarding "git log" on "git series" metadata

[Jacob Keller]

On Sun, Nov 6, 2016 at 9:33 AM, Josh Triplett <josh@joshtriplett.org> wrote:
> On Sun, Nov 06, 2016 at 09:14:56AM -0800, Junio C Hamano wrote:
>> Josh Triplett <josh@joshtriplett.org> writes:
>> > We could, but if we (or one of the many third-party git implementations)
>> > miss a case, gitlinks+reachability may appear to work in many cases with
>> > dataloss afterward, while gitrefs will fail early and not appear
>> > functional.
>>
>> I wonder what happens if we do not introduce the "gitref" but
>> instead change the behaviour of "gitlink" to imply an optional
>> reachability.  That is, when enumerating what is reachable in your
>> repository, if you see a gitlink and if you notice that you locally
>> have the target of that gitlink, you follow, but if you know you
>> lack it, you do not error out.  This may be making things too
>> complex to feasibily implement by simplify them ;-) and I see a few
>> immediate fallout that needs to be thought through (i.e. downsides)
>> and a few upsides, too.  I am feeling feverish and not thinking
>> straight, so I won't try to weigh pros-and-cons.
>>
>> This would definitely need protocol extension when transferring
>> objects across repositories.
>
> It'd also need a repository format extension locally.  Otherwise, if you
> ever touched that repository with an older git (or a tool built on an
> older libgit2 or JGit or other library), you could lose data.
>
> It does seem conceptually appealing, though.  In an ideal world, the
> original version of gitlink would have had opt-out reachability (and
> .gitmodules with an external repository reference could count as opting
> out).
>
> But I can't think of any case where it's OK for a git implementation to
> not know about this reachability extension and still operate on the
> gitlink.  And given that, it might as well use a new object type that
> the old version definitely won't think it understands.
>
> - Josh Triplett

That's still only true if the receiving end runs fsck, isn't it? I
suppose that's a large number of receivers, and at least there are
ways post-push to determine that objects don't make sense to that
version of git.

I think using a new mode is the safest way, and it allows easily
implementing RefTrees as well as other projects. Additionally, if we
*wanted* additional "opt-in / opt-out" support we could add this by
default to gitrefs,and they could (possibly) replace gitlinks in the
future?

Thanks,
Jake

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Sun, Nov 06, 2016 at 12:17:10PM -0800, Jacob Keller wrote:
> On Sun, Nov 6, 2016 at 9:33 AM, Josh Triplett <josh@joshtriplett.org> wrote:
> > On Sun, Nov 06, 2016 at 09:14:56AM -0800, Junio C Hamano wrote:
> >> Josh Triplett <josh@joshtriplett.org> writes:
> >> > We could, but if we (or one of the many third-party git implementations)
> >> > miss a case, gitlinks+reachability may appear to work in many cases with
> >> > dataloss afterward, while gitrefs will fail early and not appear
> >> > functional.
> >>
> >> I wonder what happens if we do not introduce the "gitref" but
> >> instead change the behaviour of "gitlink" to imply an optional
> >> reachability.  That is, when enumerating what is reachable in your
> >> repository, if you see a gitlink and if you notice that you locally
> >> have the target of that gitlink, you follow, but if you know you
> >> lack it, you do not error out.  This may be making things too
> >> complex to feasibily implement by simplify them ;-) and I see a few
> >> immediate fallout that needs to be thought through (i.e. downsides)
> >> and a few upsides, too.  I am feeling feverish and not thinking
> >> straight, so I won't try to weigh pros-and-cons.
> >>
> >> This would definitely need protocol extension when transferring
> >> objects across repositories.
> >
> > It'd also need a repository format extension locally.  Otherwise, if you
> > ever touched that repository with an older git (or a tool built on an
> > older libgit2 or JGit or other library), you could lose data.
> >
> > It does seem conceptually appealing, though.  In an ideal world, the
> > original version of gitlink would have had opt-out reachability (and
> > .gitmodules with an external repository reference could count as opting
> > out).
> >
> > But I can't think of any case where it's OK for a git implementation to
> > not know about this reachability extension and still operate on the
> > gitlink.  And given that, it might as well use a new object type that
> > the old version definitely won't think it understands.
> >
> > - Josh Triplett
> 
> That's still only true if the receiving end runs fsck, isn't it? I
> suppose that's a large number of receivers, and at least there are
> ways post-push to determine that objects don't make sense to that
> version of git.
> 
> I think using a new mode is the safest way, and it allows easily
> implementing RefTrees as well as other projects. Additionally, if we
> *wanted* additional "opt-in / opt-out" support we could add this by
> default to gitrefs,and they could (possibly) replace gitlinks in the
> future?

Once we have gitrefs, you have both alternatives: reachable (gitref) or
not reachable (gitlink).

However, if you want some way to mark reachable objects as not
reachable, such as for a sparse checkout, external large-object storage,
or similar, then you can use a single unified mechanism for that whether
working with gitrefs, trees, or blobs.

Re: Regarding "git log" on "git series" metadata

[Jacob Keller]

On Sun, Nov 6, 2016 at 5:18 PM, Josh Triplett <josh@joshtriplett.org> wrote:
> Once we have gitrefs, you have both alternatives: reachable (gitref) or
> not reachable (gitlink).
>
> However, if you want some way to mark reachable objects as not
> reachable, such as for a sparse checkout, external large-object storage,
> or similar, then you can use a single unified mechanism for that whether
> working with gitrefs, trees, or blobs.

Fair enough.

Thanks,
Jake

Re: Regarding "git log" on "git series" metadata

[Duy Nguyen]

On Mon, Nov 7, 2016 at 8:18 AM, Josh Triplett <josh@joshtriplett.org> wrote:
> Once we have gitrefs, you have both alternatives: reachable (gitref) or
> not reachable (gitlink).
>
> However, if you want some way to mark reachable objects as not
> reachable, such as for a sparse checkout, external large-object storage,
> or similar, then you can use a single unified mechanism for that whether
> working with gitrefs, trees, or blobs.

How? Whether an object reachable or not is baked in the definition (of
either gitlink or gitref). I don't think you can have a "maybe
reachable" type then rely on an external source to determine
reachability,
-- 
Duy

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Mon, Nov 07, 2016 at 04:42:04PM +0700, Duy Nguyen wrote:
> On Mon, Nov 7, 2016 at 8:18 AM, Josh Triplett <josh@joshtriplett.org> wrote:
> > Once we have gitrefs, you have both alternatives: reachable (gitref) or
> > not reachable (gitlink).
> >
> > However, if you want some way to mark reachable objects as not
> > reachable, such as for a sparse checkout, external large-object storage,
> > or similar, then you can use a single unified mechanism for that whether
> > working with gitrefs, trees, or blobs.
> 
> How? Whether an object reachable or not is baked in the definition (of
> either gitlink or gitref). I don't think you can have a "maybe
> reachable" type then rely on an external source to determine
> reachability,

You'd have various "reachable by default" entries in trees, including
trees, blobs, and gitrefs, and then have an external mechanism (likely
via .git/config) to say "ignore objects with these hashes/paths".  For
instance, you might say "ignore all objects only reachable from the path
'assets/video/*' within a commit's tree".  With the right set of client
and server extensions, you could then avoid downloading those objects.

Re: Regarding "git log" on "git series" metadata

[Junio C Hamano]

Josh Triplett <josh@joshtriplett.org> writes:

>> This would definitely need protocol extension when transferring
>> objects across repositories.
>
> It'd also need a repository format extension locally.  Otherwise, if you
> ever touched that repository with an older git (or a tool built on an
> older libgit2 or JGit or other library), you could lose data.

True.  Thanks for sanity-checking me.

Re: Regarding "git log" on "git series" metadata

[Josh Triplett]

On Fri, Nov 04, 2016 at 10:57:09AM -0700, Junio C Hamano wrote:
> After your talk at LPC2016, I was thinking about your proposal to
> give an option to hide certain parents from "git log" traversal.
> 
> While I do not think we would terribly mind a new feature in the
> core to support third-party additions like "git series" better, I
> think this particular one is a big mistake that we shouldn't take.
[...]
> I think this is backwards.  The root cause of the issue you have
> with "gitk" is because you added something that is *NOT* a parent to
> your commit.  We shouldn't have to add a mechanism to filter
> something that shouldn't have been added there in the first place.
> 
> I am wondering if an alternative approach would work better.
> 
> Imagine we invent a new tree entry type, "gitref", that is similar
> to "gitlink" in that it can record a commit object name in a tree,
> but unlike "gitlink" it does imply reachability.  And you do not add
> phony parents to your commit object.  A tree that has "gitref"s in
> it is about annotating the commits in the same repository (e.g. the
> tree references two commits, "base" and "tip", to point into a slice
> of the main history).  And it is perfectly sensible for such a
> pointer to imply reachability---after all it serves different
> purposes from "gitlink".

I absolutely agree with this, and I'd love to have gitref or similar in
core git.  Given the availability of that mechanism, I'd love to use it
in git-series.  (And in git submodule, as well, for other projects.)

The one critical issue there, though: that would break backward
compatibility with old versions of git.  No old version of git could
push, pull, gc, repack, or otherwise touch a repository that used this
feature.

The advantages of the approach (viewing and manipulating the series with
pure git) seem sufficiently high to make that worth considering, but it
is a significant downside.

> Another alternative that I am negative about (but is probably a
> better hack than how you abused the "parent" link) might be to add a
> new commit object header field that behaves similarly to "parent"
> only in that it implies reachability.  But recording the extra
> parent in commit object was not something you wanted to do in the
> first place (i.e. your series processing is done solely on the
> contents of the tree, and you do not read this extra parent). If you
> need to add an in-tree reference to another commit in your future
> versions of "git series", with either this variant or your original
> implementation, you would end up needing adding more "parent" (or
> pseudo parent) only to preserve reachability.  At that point, I
> think it makes more sense to have entries in the tree to directly
> ensure reachability, if you want these entries to always point at an
> in-tree object.

This would similarly break compatibility with old git, as old git
wouldn't follow those reachability-only links from commits, so it could
throw away the data.

One approach compatible with old git would be to continue adding the
relevant commits as artificial parents, but have a separate commit
metadata field that says which parents to ignore; old git would then do
the right thing, as long as it doesn't rewrite the commit entirely.

That does have the same disadvantages of having to duplicate the
information in both the tree and the parent list, though; it's the same
class of hack, just with improved usability.  I'd much rather use
gitrefs.

> I am afraid that I probably am two steps ahead of myself, because I
> am reasonably sure that it is quite possible that I have overlooked
> something trivially obvious that makes the "gitref" approach
> unworkable.

gitref seems like a good idea to me, as long as we can sort out the
compatibility story.