Re: Partial clone design (with connectivity check for locally-created objects)

[Jonathan Nieder]

Hi,

Ben Peart wrote:

> We've discussed a couple of different possible solutions, each of
> which have different tradeoffs.  Let me try to summarize here and
> perhaps suggest some other possibilities:

Thanks for this.  Some comments below.

> Promised list
> -
> This provides an external data structure that allowed us to flag
> objects that came from a remote server (vs created locally).
>
> The biggest drawback is that this data structure can get very large
> and become difficult/expensive to generate/transfer/maintain.

Agreed.  Using a single immutable file to maintain this data with
lock-and-rename update means that I/O when updating it can be a
bottleneck and that contention can be a problem.

> It also (at least in one proposal) required protocol and server side
> changes to support it.

I don't think that's a very big problem.  This is the Git project: we
control the protocol and the server.  Partial clone requires changing
the protocol and server already.

> Annotated via filename
> --
> This idea is to annotate the file names of objects that came from a
> remote server (pack files and loose objects) with a unique file
> extension (.remote) that indicates whether they are locally created
> or not.
>
> To make this work, git must understand about both types of loose
> objects and pack files and search in both locations when looking for
> objects.

I don't understand the drawback you're describing here.  To avoid a
number of serious problems, Git already needs to be aware of partial
clone.  I don't think anyone has been proposing adding partial clone
to upstream Git without a repository format extension (see
Documentation/technical/repository-version.txt) to prevent older
versions of Git from being confused about such repositories.

If you don't do this, some problems include
- confusing messages due to missing objects
- errors over the wire protocol from trying to serve fetches and
  getting confused
- "git gc" running and not knowing which objects are safe to be
  deleted

So the relevant issue couldn't be that Git has to be changed at all:
it would be that a change is excessively invasive.

But it's not clear to me that the change you are describing is very
invasive.

> Another drawback of this is that commands (repack, gc) that optimize
> loose objects and pack files must now be aware of the different
> extensions and handle both while not merging remote and non-remote
> objects.
>
> In short, we're creating separate object stores - one for locally
> created objects and one for everything else.

These also seem like non-issues.

Some examples of problems I could imagine:

- is writing multiple files when writing a loose object a problem for
  your setup?
- is one of the operations described (repack, prune, fsck) too slow?

Do you forsee either of those being an issue?

> Now a couple of different ideas:
>
> Annotated via flags
> ===
> The fundamental idea here is that we add the ability to flag locally
> created objects on the object itself.

Do you mean changing the underlying object format that produces an
object's object id?  Or do you mean changing the container format?

Changing the container format is exactly what was described in the
previous example ("Annotated via filename").  There are other ways to
change the container format: e.g. if writing multiple files when
writing a loose object is a problem, we could add a field that does
not affect the object id to the loose object format.

[...]
> Local list
> --
> Given the number of locally created objects is usually very small in
> comparison to the total number of objects (even just due to
> history), it makes more sense to track locally created objects
> instead of promised/remote objects.
>
> The biggest advantage of this over the "promised list" is that the
> "local list" being maintained is _significantly_ smaller (often
> orders of magnitude smaller).
[...]
> On the surface, this seems like the simplest solution that meets the
> stated requirements.

This has the same problems as the list of promised objects: excessive
I/O and contention when updating the list.

Moreover, it doesn't bring one of the main benefits of the list of
promised objects.  Promised objects are not present in the local
repository, so the list of promises provided a way to maintain some
information about them (e.g., object size).  Locally created objects
are present in the local repository so they don't need such metadata.

> Object DB
> -

If I understand correctly, this is pushing the issues described in the
other cases into a hook and making them not upstream Git's problem.

But it is still someone's problem.  It just means upstream Git doesn't
benefit from their solution to it.

I don't see a need to give up in that way just yet.

I'm also available on #git-devel on freenode.net for real-time
conversation.  Logs are at http://bit.ly/aLzrmv.  You can prepend a
message with "[off]" to prevent 

Re: Partial clone design (with connectivity check for locally-created objects)

[Ben Peart]

On 8/7/2017 3:41 PM, Junio C Hamano wrote:

Ben Peart  writes:


My concern with this proposal is the combination of 1) writing a new
pack file for every git command that ends up bringing down a missing
object and 2) gc not compressing those pack files into a single pack
file.


Your noticing these is a sign that you read the outline of the
design correctly, I think.

The basic idea is that the local fsck should tolerate missing
objects when they are known to be obtainable from that external
service, but should still be able to diagnose missing objects that
we do not know if the external service has, especially the ones that
have been newly created locally and not yet made available to them
by pushing them back.



This helps me a lot as now I think I understand the primary requirement 
we're trying to solve for.  Let me rephrase it and see if this makes sense:


We need to be able to identify whether an object was created locally 
(and should pass more strict fsck/connectivity tests) or whether it came 
from a remote (and so any missing objects could presumably be fetched 
from the server).


I agree it would be nice to solve this (and not just punt fsck - even if 
it is an opt-in behavior).


We've discussed a couple of different possible solutions, each of which 
have different tradeoffs.  Let me try to summarize here and perhaps 
suggest some other possibilities:



Promised list
-
This provides an external data structure that allowed us to flag objects 
that came from a remote server (vs created locally).


The biggest drawback is that this data structure can get very large and 
become difficult/expensive to generate/transfer/maintain.


It also (at least in one proposal) required protocol and server side 
changes to support it.



Annotated via filename
--
This idea is to annotate the file names of objects that came from a 
remote server (pack files and loose objects) with a unique file 
extension (.remote) that indicates whether they are locally created or not.


To make this work, git must understand about both types of loose objects 
and pack files and search in both locations when looking for objects.


Another drawback of this is that commands (repack, gc) that optimize 
loose objects and pack files must now be aware of the different 
extensions and handle both while not merging remote and non-remote objects.


In short, we're creating separate object stores - one for locally 
created objects and one for everything else.



Now a couple of different ideas:

Annotated via flags
===
The fundamental idea here is that we add the ability to flag locally 
created objects on the object itself.


Given that at the core, "Git is a simple key-value data store" can we 
take advantage of that fact and include a "locally created" bit as a 
property on every object?


I could not think of a good way to accomplish this as it is ultimately 
changing the object format which creates rapidly expanding ripples of 
change.


For example, The object header currently includes the type a space, the 
length and a null. Even if we could add a "local" property (either by 
adding a 5th item, taking over the space, creating new object types, 
etc), the fact that the header is included in the sha1 means that push 
would become problematic as flipping the bit would change the sha and 
the trees and commits that reference it.



Local list
--
Given the number of locally created objects is usually very small in 
comparison to the total number of objects (even just due to history), it 
makes more sense to track locally created objects instead of 
promised/remote objects.


The biggest advantage of this over the "promised list" is that the 
"local list" being maintained is _significantly_ smaller (often orders 
of magnitude smaller).


Another advantage over the "promised list" solution is that it doesn't 
require any server side or protocol changes.


On the client when objects are created (write_loose_object?) the new 
objects are added to the "local list" and in the connectivity check 
(fsck) if the object is not in the "local list," the connectivity check 
can be skipped as any missing object can presumably be retrieved from 
the server.


A simple file format could be used (header + list of SHA1 values) and 
write_loose_object could do a trivial append. In fsck, the file could be 
loaded into a hashmap to make for fast existence checks.


Entries could be removed from the "local list" for objects later fetched 
from a server (though I had a hard time contriving a scenario where this 
would happen so I consider this optional).


On the surface, this seems like the simplest solution that meets the 
stated requirements.



Object DB
-
This is a different way of providing separate object stores than the 
"Annotated via filename" proposal. It should be a cleaner/more elegant 
solution that enables several other capabilities but it is also more 

Re: Partial clone design (with connectivity check for locally-created objects)

[Ben Peart]

On 8/7/2017 3:21 PM, Jonathan Nieder wrote:

Hi,

Ben Peart wrote:

On Fri, 04 Aug 2017 15:51:08 -0700
Junio C Hamano  wrote:

Jonathan Tan  writes:



"Imported" objects must be in a packfile that has a ".remote"
file with arbitrary text (similar to the ".keep" file). They come from
clones, fetches, and the object loader (see below).
...

A "homegrown" object is valid if each object it references:
  1. is a "homegrown" object,
  2. is an "imported" object, or
  3. is referenced by an "imported" object.


Overall it captures what was discussed, and I think it is a good
start.


I missed the offline discussion and so am trying to piece together
what this latest design is trying to do.  Please let me know if I'm
not understanding something correctly.


I believe
https://public-inbox.org/git/cover.1501532294.git.jonathanta...@google.com/
and the surrounding thread (especially
https://public-inbox.org/git/xmqqefsudjqk@gitster.mtv.corp.google.com/)
is the discussion Junio is referring to.

[...]

This segmentation is what is driving the need for the object loader
to build a new local pack file for every command that has to fetch a
missing object.  For example, we can't just write a tree object from
a "partial" clone into the loose object store as we have no way for
fsck to treat them differently and ignore any missing objects
referenced by that tree object.


That's related and how it got lumped into this proposal, but it's not
the only motivation.

Other aspects:

  1. using pack files instead of loose objects means we can use deltas.
 This is the primary motivation.

  2. pack files can use reachability bitmaps (I realize there are
 obstacles to getting benefit out of this because git's bitmap
 format currently requires a pack to be self-contained, but I
 thought it was worth mentioning for completeness).

  3. existing git servers are oriented around pack files; they can
 more cheaply serve objects from pack files in pack format,
 including reusing deltas from them.

  4. file systems cope better with a few large files than many small
 files

[...]

We all know that git doesn't scale well with a lot of pack files as
it has to do a linear search through all the pack files when
attempting to find an object.  I can see that very quickly, there
would be a lot of pack files generated and with gc ignoring
"partial" pack files, this would never get corrected.


Yes, that's an important point.  Regardless of this proposal, we need
to get more aggressive about concatenating pack files (e.g. by
implementing exponential rollup in "git gc --auto").


In our usage scenarios, _all_ of the objects come from "partial"
clones so all of our objects would end up in a series of "partial"
pack files and would have pretty poor performance as a result.


Can you say more about this?  Why would the pack files (or loose
objects, for that matter) never end up being consolidated into few
pack files?



Our initial clone is very sparse - we only pull down the commit we are 
about to checkout and none of the blobs. All missing objects are then 
downloaded on demand (and in this proposal, would end up in a "partial" 
pack file).  For performance reasons, we also (by default) download a 
server computed pack file of commits and trees to pre-populate the local 
cache.


Without modification, fsck, repack, prune, gc will trigger every object 
in the repo to be downloaded.  We punted for now and just block those 
commands but eventually they need to be aware of missing objects so that 
they do not cause them to be downloaded.  Jonathan is already working on 
this for fsck in another patch series.



[...]

That thinking did lead me back to wondering again if we could live
with a repo specific flag.  If any clone/fetch was "partial" the
flag is set and fsck ignore missing objects whether they came from a
"partial" remote or not.

I'll admit it isn't as robust if someone is mixing and matching
remotes from different servers some of which are partial and some of
which are not.  I'm not sure how often that would actually happen
but I _am_ certain a single repo specific flag is a _much_ simpler
model than anything else we've come up with so far.


The primary motivation in this thread is locally-created objects, not
objects obtained from other remotes.  Objects obtained from other
remotes are more of an edge case.



Thank you - that helps me to better understand the requirements of the 
problem we're trying to solve.  In short, that means what we really need 
is a way to identify locally created objects so that fsck can do a 
complete connectivity check on them.  I'll have to think about a good 
way to do that - we've talked about a few but each has a different set 
of trade-offs and none of them are great (yet :)).



Thanks for your thoughtful comments.

Jonathan

Re: Partial clone design (with connectivity check for locally-created objects)

[Jonathan Tan]

On Mon, 7 Aug 2017 15:12:11 -0400
Ben Peart  wrote:

> I missed the offline discussion and so am trying to piece together what 
> this latest design is trying to do.  Please let me know if I'm not 
> understanding something correctly.
> 
>  From what I can tell, objects are going to be segmented into two 
> "types" - those that were fetched from a remote source that allows 
> partial clones/fetches (lazyobject/imported) and those that come from 
> "regular" remote sources (homegrown) that requires all objects to exist 
> locally.
> 
> FWIW, the names here are not making things clearer for me. If I'm 
> correct perhaps "partial" and "normal" would be better to indicate the 
> type of the source? Anyway...

That's right. As for names, I'm leaning now towards "imported" and
"non-imported". "Partial" is a bit strange because such an object is
fully available; it's just that the objects that it references are
promised by the server.

> Once the objects are segmented into the 2 types, the fsck connectivity 
> check code is updated to ignore missing objects from "partial" remotes 
> but still expect/validate them from "normal" remotes.
> 
> This compromise seems reasonable - don't generate errors for missing 
> objects for remotes that returned a partial clone but do generate errors 
> for missing objects from normal clones as a missing object is always an 
> error in this case.

Yes. In addition, the references of "imported" objects are also
potentially used when connectivity-checking "non-imported" objects - if
a "non-imported" object refers to an object that an "imported" object
refers to, that is fine, even though we don't have that object.

> This segmentation is what is driving the need for the object loader to 
> build a new local pack file for every command that has to fetch a 
> missing object.  For example, we can't just write a tree object from a 
> "partial" clone into the loose object store as we have no way for fsck 
> to treat them differently and ignore any missing objects referenced by 
> that tree object.
> 
> My concern with this proposal is the combination of 1) writing a new 
> pack file for every git command that ends up bringing down a missing 
> object and 2) gc not compressing those pack files into a single pack file.
> 
> We all know that git doesn't scale well with a lot of pack files as it 
> has to do a linear search through all the pack files when attempting to 
> find an object.  I can see that very quickly, there would be a lot of 
> pack files generated and with gc ignoring "partial" pack files, this 
> would never get corrected.
> 
> In our usage scenarios, _all_ of the objects come from "partial" clones 
> so all of our objects would end up in a series of "partial" pack files 
> and would have pretty poor performance as a result.

One possible solution...would support for annotating loose objects with
".remote" be sufficient? (That is, for each loose object file created,
create another of the same name but with ".remote" appended.) This means
that a loose-object-creating lazy loader would need to create 2 files
per object instead of one.

The lazy loader protocol will thus be updated to something resembling a
prior version with the loader writing objects directly to the object
database, but now the loader is also responsible for creating the
".remote" files.  (In the Android use case, we probably won't need the
writing-to-partial-packfile mechanism anymore since only comparatively
few and large blobs will go in there.)

Re: Partial clone design (with connectivity check for locally-created objects)

[Junio C Hamano]

Ben Peart  writes:

> My concern with this proposal is the combination of 1) writing a new
> pack file for every git command that ends up bringing down a missing
> object and 2) gc not compressing those pack files into a single pack
> file.

Your noticing these is a sign that you read the outline of the
design correctly, I think.  

The basic idea is that the local fsck should tolerate missing
objects when they are known to be obtainable from that external
service, but should still be able to diagnose missing objects that
we do not know if the external service has, especially the ones that
have been newly created locally and not yet made available to them
by pushing them back.

So we need a way to tell if an object that we do not have (but we
know about) can later be obtained from the external service.
Maintaining an explicit list of such objects obviously is one way,
but we can get the moral equivalent by using pack files.  After
receiving a pack file that has a commit from such an external
service, if the commit refers to its parent commit that we do not
have locally, the design proposes us to consider that the parent
commit that is missing is available at the external service that
gave the pack to us.  Similarly for missing trees, blobs, and any
objects that are supposed to be "reachable" from objects in such a
packfile.  

We can extend the approach to cover loose objects if we wanted to;
just define an alternate object store used internally for this
purpose and drop loose objects obtained from such an external
service in that object store.

Because we do not want to leave too many loose objects and small
packfiles lying around, we will need a new way of packing these.
Just enumerate these objects known to have come from the external
service (by being in packfiles marked as such or being loose objects
in the dedicated alternate object store), and create a single larger
packfile, which is marked as "holding the objects that are known to
be in the external service".  We do not have such a mode of gc, and
that is a new development that needs to happen, but we know that is
doable.

> That thinking did lead me back to wondering again if we could live
> with a repo specific flag.  If any clone/fetch was "partial" the flag
> is set and fsck ignore missing objects whether they came from a
> "partial" remote or not.

The only reason people run "git fsck" is to make sure that their
local repository is sound and they can rely on the objects you have
as the base of building new stuff on top of.  That is why we are
trying to find a way to make sure "fsck" can be used to detect
broken or missing objects that cannot be obtained from the
lazy-object store, without incurring undue overhead for normal
codepath (i.e. outside fsck).

It is OK to go back to wondering again, but I think that essentially
tosses "git fsck" out of the window and declares that it is OK to
hope that local objects will never go bad.  We can make such an
declaration anytime, but I do not want to see us doing so without
first trying to solve the issue without punting.

Re: Partial clone design (with connectivity check for locally-created objects)

[Jonathan Nieder]

Hi,

Ben Peart wrote:
>> On Fri, 04 Aug 2017 15:51:08 -0700
>> Junio C Hamano  wrote:
>>> Jonathan Tan  writes:

 "Imported" objects must be in a packfile that has a ".remote"
 file with arbitrary text (similar to the ".keep" file). They come from
 clones, fetches, and the object loader (see below).
 ...

 A "homegrown" object is valid if each object it references:
  1. is a "homegrown" object,
  2. is an "imported" object, or
  3. is referenced by an "imported" object.
>>>
>>> Overall it captures what was discussed, and I think it is a good
>>> start.
>
> I missed the offline discussion and so am trying to piece together
> what this latest design is trying to do.  Please let me know if I'm
> not understanding something correctly.

I believe
https://public-inbox.org/git/cover.1501532294.git.jonathanta...@google.com/
and the surrounding thread (especially
https://public-inbox.org/git/xmqqefsudjqk@gitster.mtv.corp.google.com/)
is the discussion Junio is referring to.

[...]
> This segmentation is what is driving the need for the object loader
> to build a new local pack file for every command that has to fetch a
> missing object.  For example, we can't just write a tree object from
> a "partial" clone into the loose object store as we have no way for
> fsck to treat them differently and ignore any missing objects
> referenced by that tree object.

That's related and how it got lumped into this proposal, but it's not
the only motivation.

Other aspects:

 1. using pack files instead of loose objects means we can use deltas.
This is the primary motivation.

 2. pack files can use reachability bitmaps (I realize there are
obstacles to getting benefit out of this because git's bitmap
format currently requires a pack to be self-contained, but I
thought it was worth mentioning for completeness).

 3. existing git servers are oriented around pack files; they can
more cheaply serve objects from pack files in pack format,
including reusing deltas from them.

 4. file systems cope better with a few large files than many small
files

[...]
> We all know that git doesn't scale well with a lot of pack files as
> it has to do a linear search through all the pack files when
> attempting to find an object.  I can see that very quickly, there
> would be a lot of pack files generated and with gc ignoring
> "partial" pack files, this would never get corrected.

Yes, that's an important point.  Regardless of this proposal, we need
to get more aggressive about concatenating pack files (e.g. by
implementing exponential rollup in "git gc --auto").

> In our usage scenarios, _all_ of the objects come from "partial"
> clones so all of our objects would end up in a series of "partial"
> pack files and would have pretty poor performance as a result.

Can you say more about this?  Why would the pack files (or loose
objects, for that matter) never end up being consolidated into few
pack files?

[...]
> That thinking did lead me back to wondering again if we could live
> with a repo specific flag.  If any clone/fetch was "partial" the
> flag is set and fsck ignore missing objects whether they came from a
> "partial" remote or not.
>
> I'll admit it isn't as robust if someone is mixing and matching
> remotes from different servers some of which are partial and some of
> which are not.  I'm not sure how often that would actually happen
> but I _am_ certain a single repo specific flag is a _much_ simpler
> model than anything else we've come up with so far.

The primary motivation in this thread is locally-created objects, not
objects obtained from other remotes.  Objects obtained from other
remotes are more of an edge case.

Thanks for your thoughtful comments.

Jonathan

Re: Partial clone design (with connectivity check for locally-created objects)

[Ben Peart]

On 8/4/2017 8:21 PM, Jonathan Tan wrote:

On Fri, 04 Aug 2017 15:51:08 -0700
Junio C Hamano  wrote:


Jonathan Tan  writes:


"Imported" objects must be in a packfile that has a ".remote"
file with arbitrary text (similar to the ".keep" file). They come from
clones, fetches, and the object loader (see below).
...
A "homegrown" object is valid if each object it references:
  1. is a "homegrown" object,
  2. is an "imported" object, or
  3. is referenced by an "imported" object.


Overall it captures what was discussed, and I think it is a good
start.


I missed the offline discussion and so am trying to piece together what 
this latest design is trying to do.  Please let me know if I'm not 
understanding something correctly.


From what I can tell, objects are going to be segmented into two 
"types" - those that were fetched from a remote source that allows 
partial clones/fetches (lazyobject/imported) and those that come from 
"regular" remote sources (homegrown) that requires all objects to exist 
locally.


FWIW, the names here are not making things clearer for me. If I'm 
correct perhaps "partial" and "normal" would be better to indicate the 
type of the source? Anyway...


Once the objects are segmented into the 2 types, the fsck connectivity 
check code is updated to ignore missing objects from "partial" remotes 
but still expect/validate them from "normal" remotes.


This compromise seems reasonable - don't generate errors for missing 
objects for remotes that returned a partial clone but do generate errors 
for missing objects from normal clones as a missing object is always an 
error in this case.


This segmentation is what is driving the need for the object loader to 
build a new local pack file for every command that has to fetch a 
missing object.  For example, we can't just write a tree object from a 
"partial" clone into the loose object store as we have no way for fsck 
to treat them differently and ignore any missing objects referenced by 
that tree object.


My concern with this proposal is the combination of 1) writing a new 
pack file for every git command that ends up bringing down a missing 
object and 2) gc not compressing those pack files into a single pack file.


We all know that git doesn't scale well with a lot of pack files as it 
has to do a linear search through all the pack files when attempting to 
find an object.  I can see that very quickly, there would be a lot of 
pack files generated and with gc ignoring "partial" pack files, this 
would never get corrected.


In our usage scenarios, _all_ of the objects come from "partial" clones 
so all of our objects would end up in a series of "partial" pack files 
and would have pretty poor performance as a result.


I wondered if it is possible to flag a specific remote as "partial" and 
have fsck be able to track any given object back to the remote and then 
properly handle the fact that it was missing based on that. I couldn't 
think of a good way to do that without some additional data structure 
that would have to be build/maintained (ie promises).


That thinking did lead me back to wondering again if we could live with 
a repo specific flag.  If any clone/fetch was "partial" the flag is set 
and fsck ignore missing objects whether they came from a "partial" 
remote or not.


I'll admit it isn't as robust if someone is mixing and matching remotes 
from different servers some of which are partial and some of which are 
not.  I'm not sure how often that would actually happen but I _am_ 
certain a single repo specific flag is a _much_ simpler model than 
anything else we've come up with so far.




I doubt you want to treat all fetches/clones the same way as the
"lazy object" loading, though.  You may be critically rely on the
corporate central server that will give the objects it "promised"
when you cloned from it lazily (i.e. it may have given you a commit,
but not its parents or objects contained in its tree--you still know
that the parents and the tree and its contents will later be
available and rely on that fact).  You trust that and build on top,
so the packfile you obtained when you cloned from such a server
should count as "imported".  But if you exchanged wip changes with
your colleages by fetching or pushing peer-to-peer, without the
corporate central server knowing, you would want to treat objects in
packs (or loose objects) you obtained that way as "not imported".


That's true. I discussed this with a teammate and we might need to make
extensions.lazyObject be the name of the "corporate central server"
remote instead, and have a "loader" setting within that remote, so that
we can distinguish that objects from this server are "imported" but
objects from other servers are not.

The connectivity check shouldn't be slow in this case because fetches
are usually onto tips that we have (so we don't hit case 3).


Also I think "imported" vs "homegrown" may be a bit misnomer; 

Re: Partial clone design (with connectivity check for locally-created objects)

[Jonathan Tan]

On Fri, 04 Aug 2017 15:51:08 -0700
Junio C Hamano  wrote:

> Jonathan Tan  writes:
> 
> > "Imported" objects must be in a packfile that has a ".remote"
> > file with arbitrary text (similar to the ".keep" file). They come from
> > clones, fetches, and the object loader (see below).
> > ...
> > A "homegrown" object is valid if each object it references:
> >  1. is a "homegrown" object,
> >  2. is an "imported" object, or
> >  3. is referenced by an "imported" object.
> 
> Overall it captures what was discussed, and I think it is a good
> start.
> 
> I doubt you want to treat all fetches/clones the same way as the
> "lazy object" loading, though.  You may be critically rely on the
> corporate central server that will give the objects it "promised"
> when you cloned from it lazily (i.e. it may have given you a commit,
> but not its parents or objects contained in its tree--you still know
> that the parents and the tree and its contents will later be
> available and rely on that fact).  You trust that and build on top,
> so the packfile you obtained when you cloned from such a server
> should count as "imported".  But if you exchanged wip changes with
> your colleages by fetching or pushing peer-to-peer, without the
> corporate central server knowing, you would want to treat objects in
> packs (or loose objects) you obtained that way as "not imported".

That's true. I discussed this with a teammate and we might need to make
extensions.lazyObject be the name of the "corporate central server"
remote instead, and have a "loader" setting within that remote, so that
we can distinguish that objects from this server are "imported" but
objects from other servers are not.

The connectivity check shouldn't be slow in this case because fetches
are usually onto tips that we have (so we don't hit case 3).

> Also I think "imported" vs "homegrown" may be a bit misnomer; the
> idea to split objects into two camps sounds like a good idea, and
> "imported" probably is an OK name to use for the category that is a
> group of objects to which you know/trust are backed by your lazy
> loader.  But the other one does not have to be "home"-grown.
> 
> Well, the names are not that important, but I think the line between
> the two classes should not be "everything that came from clone and
> fetch is imported", which is a more important point I am trying to
> make.
> 
> Thanks.

Maybe "imported" vs "non-imported" would be better. I agree that the
objects in the non-"imported" group could still be obtained from
elsewhere.

Thanks for your comments.

Re: Partial clone design (with connectivity check for locally-created objects)

[Junio C Hamano]

Jonathan Tan  writes:

> "Imported" objects must be in a packfile that has a ".remote"
> file with arbitrary text (similar to the ".keep" file). They come from
> clones, fetches, and the object loader (see below).
> ...
> A "homegrown" object is valid if each object it references:
>  1. is a "homegrown" object,
>  2. is an "imported" object, or
>  3. is referenced by an "imported" object.

Overall it captures what was discussed, and I think it is a good
start.

I doubt you want to treat all fetches/clones the same way as the
"lazy object" loading, though.  You may be critically rely on the
corporate central server that will give the objects it "promised"
when you cloned from it lazily (i.e. it may have given you a commit,
but not its parents or objects contained in its tree--you still know
that the parents and the tree and its contents will later be
available and rely on that fact).  You trust that and build on top,
so the packfile you obtained when you cloned from such a server
should count as "imported".  But if you exchanged wip changes with
your colleages by fetching or pushing peer-to-peer, without the
corporate central server knowing, you would want to treat objects in
packs (or loose objects) you obtained that way as "not imported".

Also I think "imported" vs "homegrown" may be a bit misnomer; the
idea to split objects into two camps sounds like a good idea, and
"imported" probably is an OK name to use for the category that is a
group of objects to which you know/trust are backed by your lazy
loader.  But the other one does not have to be "home"-grown.

Well, the names are not that important, but I think the line between
the two classes should not be "everything that came from clone and
fetch is imported", which is a more important point I am trying to
make.

Thanks.

Partial clone design (with connectivity check for locally-created objects)

[Jonathan Tan]

After some discussion in [1] (in particular, about preserving the
functionality of the connectivity check as much as possible) and some
in-office discussion, here's an updated design.

Overview


This is an update of the design in [1].

The main difference between this and other related work [1] [2] [3] is
that we can still check connectivity between locally-created objects
without having to consult a remote server for any information.

In addition, the object loader writes to an incomplete packfile. This
(i) ensures that Git has immediate access to the object, (ii) ensures
that not too many files are written during a single Git invocation, and
(iii) prevents some unnecessary copies (compared to, for example,
transmitting entire objects through the protocol).

Local repo layout
=

Objects in the local repo are further divided into "homegrown" and
"imported" objects.

"Imported" objects must be in a packfile that has a ".remote"
file with arbitrary text (similar to the ".keep" file). They come from
clones, fetches, and the object loader (see below).

"Homegrown" objects are every other object.

Object loader
=

The object loader is a process that can obtain objects from elsewhere,
given their hashes, and write their packed representation to a
client-given file.

The first time a missing object is needed during an invocation of Git,
Git creates a temporary packfile and writes the header with a
placeholder number of objects. Then, it starts the object loader,
passing in the name of that temporary packfile.

Whenever a missing object is needed, Git sends the hash of the missing
object and expects the loader to append (with O_APPEND) the object to
that packfile. Git keeps track of the object offsets as it goes, and Git
can use the contents of that incomplete packfile. This is similar to
what "git fast-import" does.

When Git exits, it writes the number of objects in the header, writes
the packfile checksum, moves the packfile to its final location, and
writes a .idx and a .remote file.

Connectivity check
==

An object walk is performed as usual from the tips (see the
documentation for fsck etc. for which tips they use).

A "homegrown" object is valid if each object it references:
 1. is a "homegrown" object,
 2. is an "imported" object, or
 3. is referenced by an "imported" object.

The references of an "imported" object are not checked.

Performance notes
-

Because of rule 3 above, iteration through every "imported" object (or,
at least, every "imported" object of a certain type) is sometimes
required.

For fsck, this should be fine because (i) this is not a regression since
currently all objects must be iterated through anyway, and (ii) fsck
prioritizes correctness over speed.

For fetch, the speed of the connectivity check is immaterial; the
connectivity check no longer needs to be performed because all objects
obtained from the remote are, by definition, "imported" objects.

There might be connectivity checks run during other commands like
"receive-pack". I don't expect partial clones to use these often. These
commands will still work, but performance of these is a secondary
concern in this design.

Impact on other tools
=

"git gc" will need to not do anything to an "imported" object, even if
it is unreachable, without ensuring that the connectivity check will
succeed in that object's absence. (Special attention to rule 3 under
"Connectivity check".)

If this design stands, the initial patch set will probably have "git gc"
not touch "imported" packs at all, trivially satisfying the above. In
the future, "git gc" will either need to expel such objects into loose
objects (like what is currently done for normal packs), treating them
like a "homegrown" object (unreachable, so it won't interfere with
future connectivity checks), or delete them outright - but there may be
race conditions to think of.

"git repack" will need to differentiate between packs with ".remote" and
packs without.

[1] https://public-inbox.org/git/cover.1501532294.git.jonathanta...@google.com/
[2] https://public-inbox.org/git/20170714132651.170708-1-benpe...@microsoft.com/
[3] https://public-inbox.org/git/20170803091926.1755-1-chrisc...@tuxfamily.org/