[PATCH 0/2] thin-pack capability for send-pack/receive-pack

[PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Carlos Martín Nieto]

Hi all,

This comes as a result of the discussion starting at [0] about
git-push assuming that a server will always support thin packs. Most
out there in fact do, but this isn't necessarily the case.

Some implementations may not have support for it yet, or the server
might be running in an environment where it is not feasible for it to
try to fill in the missing objects.

Jonathan and Duy mentioned that separate patches for receive-pack and
send-pack could let us work around adding this at such a late stage,
so here they are. The second patch can maybe lie in waiting for a
while.


[0] http://thread.gmane.org/gmane.comp.version-control.git/235766/focus=236402

Carlos Martín Nieto (2):
  receive-pack: advertise thin-pack
  send-pack: only send a thin pack if the server supports it

 Documentation/technical/protocol-capabilities.txt | 20 +++++++++++++++-----
 builtin/receive-pack.c                            |  2 +-
 send-pack.c                                       |  2 ++
 3 files changed, 18 insertions(+), 6 deletions(-)

-- 
1.8.4.652.g0d6e0ce

[PATCH 1/2] receive-pack: advertise thin-pack

[Carlos Martín Nieto]

upload-pack has long advertised thin-pack, letting the clients request
these smaller packs. The client however unconditionally assumes that a
server is able to fix thin packs and there is no way of telling the
client that this is in fact not the case.

Make receive-pack advertise 'thin-pack' in anticipation of the client
toggling the assumption and document this capability when used by
receive-pack.

Signed-off-by: Carlos Martín Nieto <cmn@elego.de>
---
 Documentation/technical/protocol-capabilities.txt | 20 +++++++++++++++-----
 builtin/receive-pack.c                            |  2 +-
 2 files changed, 16 insertions(+), 6 deletions(-)

diff --git a/Documentation/technical/protocol-capabilities.txt b/Documentation/technical/protocol-capabilities.txt
index fd8ffa5..4e96d51 100644
--- a/Documentation/technical/protocol-capabilities.txt
+++ b/Documentation/technical/protocol-capabilities.txt
@@ -72,15 +72,25 @@ interleaved with S-R-Q.
 thin-pack
 ---------
 
-This capability means that the server can send a 'thin' pack, a pack
-which does not contain base objects; if those base objects are available
-on client side. Client requests 'thin-pack' capability when it
-understands how to "thicken" it by adding required delta bases making
-it self-contained.
+A thin pack is one with deltas which reference base objects not
+contained within the pack (but are known to exist at the receiving
+end). This can reduce the network traffic significantly, but it
+requires the receiving end to know how to "thicken" these packs by
+adding the missing bases to the pack.
+
+The upload-pack server advertises 'thin-pack' when it can generate and
+send a thin pack. The receive-pack server advertises 'thin-pack' when
+it knows how to "thicken" the pack it receives.
+
+Likewise, the client requests the 'thin-pack' capability when it
+understands how to "thicken" it.
 
 Client MUST NOT request 'thin-pack' capability if it cannot turn a thin
 pack into a self-contained pack.
 
+Client MUST NOT send a thin pack if the server does not advertise this
+capability.
+
 
 side-band, side-band-64k
 ------------------------
diff --git a/builtin/receive-pack.c b/builtin/receive-pack.c
index e3eb5fc..0e35c02 100644
--- a/builtin/receive-pack.c
+++ b/builtin/receive-pack.c
@@ -132,7 +132,7 @@ static void show_ref(const char *path, const unsigned char *sha1)
 	else
 		packet_write(1, "%s %s%c%s%s agent=%s\n",
 			     sha1_to_hex(sha1), path, 0,
-			     " report-status delete-refs side-band-64k quiet",
+			     " report-status delete-refs side-band-64k quiet thin-pack",
 			     prefer_ofs_delta ? " ofs-delta" : "",
 			     git_user_agent_sanitized());
 	sent_capabilities = 1;
-- 
1.8.4.652.g0d6e0ce

[PATCH 2/2] send-pack: only send a thin pack if the server supports it

[Carlos Martín Nieto]

In combination a the previous patch making receive-pack advertise the
thin-pack capability, this allows git to push to a server in a
constrained environment which is not able to fix thin packs while taking
advantage of the feature for servers which can.

Signed-off-by: Carlos Martín Nieto <cmn@elego.de>
---
 send-pack.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/send-pack.c b/send-pack.c
index 7d172ef..7b88ac8 100644
--- a/send-pack.c
+++ b/send-pack.c
@@ -205,6 +205,8 @@ int send_pack(struct send_pack_args *args,
 		quiet_supported = 1;
 	if (server_supports("agent"))
 		agent_supported = 1;
+	if (!server_supports("thin-pack"))
+		args->use_thin_pack = 0;
 
 	if (!remote_refs) {
 		fprintf(stderr, "No refs in common and none specified; doing nothing.\n"
-- 
1.8.4.652.g0d6e0ce

Re: [PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Junio C Hamano]

Carlos Martín Nieto <cmn@elego.de> writes:

> Hi all,
>
> This comes as a result of the discussion starting at [0] about
> git-push assuming that a server will always support thin packs. Most
> out there in fact do, but this isn't necessarily the case.
>
> Some implementations may not have support for it yet, or the server
> might be running in an environment where it is not feasible for it to
> try to fill in the missing objects.
>
> Jonathan and Duy mentioned that separate patches for receive-pack and
> send-pack could let us work around adding this at such a late stage,
> so here they are. The second patch can maybe lie in waiting for a
> while.

I'll queue these for now, but I doubt the wisdom of this series,
given that the ship has already sailed long time ago.

Currently, no third-party implementation of a receiving end can
accept thin push, because "thin push" is not a capability that needs
to be checked by the current clients.  People will have to wait
until the clients with 2/2 patch are widely deployed before starting
to use such a receiving end that is incapable of "thin push".

Wouldn't the world be a better place if instead they used that time
waiting to help such a third-party receiving end to implement "thin
push" support?

Re: [PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Carlos Martín Nieto]

On Wed, 2013-11-06 at 12:32 -0800, Junio C Hamano wrote:
> I'll queue these for now, but I doubt the wisdom of this series,
> given that the ship has already sailed long time ago.
> 
> Currently, no third-party implementation of a receiving end can
> accept thin push, because "thin push" is not a capability that needs
> to be checked by the current clients.  People will have to wait
> until the clients with 2/2 patch are widely deployed before starting
> to use such a receiving end that is incapable of "thin push".
> 
> Wouldn't the world be a better place if instead they used that time
> waiting to help such a third-party receiving end to implement "thin
> push" support?
> 

Support in the code isn't always enough. The particular case that
brought this on is one where the index-pack implementation can deal with
thin packs just fine.

This particular service takes the pack which the client sent and does
post-processing on it to store it elsewhere. During the receive-pack
equivalent, there is no git object db that it can query for the missing
base objects. I realise this is pretty a unusual situation.

Cheers,
   cmn

Re: [PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Junio C Hamano]

Carlos Martín Nieto <cmn@elego.de> writes:

> On Wed, 2013-11-06 at 12:32 -0800, Junio C Hamano wrote:
>> I'll queue these for now, but I doubt the wisdom of this series,
>> given that the ship has already sailed long time ago.
>> 
>> Currently, no third-party implementation of a receiving end can
>> accept thin push, because "thin push" is not a capability that needs
>> to be checked by the current clients.  People will have to wait
>> until the clients with 2/2 patch are widely deployed before starting
>> to use such a receiving end that is incapable of "thin push".
>> 
>> Wouldn't the world be a better place if instead they used that time
>> waiting to help such a third-party receiving end to implement "thin
>> push" support?
>> 
>
> Support in the code isn't always enough. The particular case that
> brought this on is one where the index-pack implementation can deal with
> thin packs just fine.
>
> This particular service takes the pack which the client sent and does
> post-processing on it to store it elsewhere. During the receive-pack
> equivalent, there is no git object db that it can query for the missing
> base objects. I realise this is pretty a unusual situation.

OK, I agree that it sounds quite niche-y, but it still is sensible.
If a receiving end does not want to (this includes "it is incapable
of doing so", but does not have to be limited to) complete a thin
pack, the series will give it such an option in the longer term.

Thanks.

Re: [PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Jeff King]

On Wed, Nov 06, 2013 at 02:25:50PM -0800, Junio C Hamano wrote:

> > Support in the code isn't always enough. The particular case that
> > brought this on is one where the index-pack implementation can deal with
> > thin packs just fine.
> >
> > This particular service takes the pack which the client sent and does
> > post-processing on it to store it elsewhere. During the receive-pack
> > equivalent, there is no git object db that it can query for the missing
> > base objects. I realise this is pretty a unusual situation.
> 
> OK, I agree that it sounds quite niche-y, but it still is sensible.
> If a receiving end does not want to (this includes "it is incapable
> of doing so", but does not have to be limited to) complete a thin
> pack, the series will give it such an option in the longer term.

I wonder if we want to make the flag go in the opposite direction, then.

Right now we have no flag, and we assume the other side can handle a
thin pack. If we add a "thin" flag, then the timeline is roughly:

  1. Receive-pack starts advertising "thin".

  2. Send-pack cannot assume lack of "thin" means the other side cannot
     handle "thin" (it might just be an older receive-pack), and keeps
     sending thin packs.

  [time passes]

  3. Send-pack can safely assume that every server has learned "thin"
     and can assume that lack of "thin" means the server does not want a
     thin pack.

In other words, the benefit happens at step 3, and we do not get any
effect until some long assumption time passes.

If we instead introduced "no-thin", it is more like:

  1. Receive-pack starts advertising "no-thin" (as dictated by
     circumstances, as Carlos describes).

  2. Send-pack which does not understand no-thin will ignore it and send
     a thin pack. This is the same as now, and the same as step 2 above.

  3. An upgraded send-pack will understand no-thin and do as the server
     asks.

So an upgraded client and server can start cooperating immediately, and
we do not have to wait for the long assumption time to pass before
applying the second half.

It is tempting to think about a "thin" flag because that would be the
natural way to have implemented it from the very beginning. But it is
not the beginning, and the negative flag is the only way at this point
to say "if you understand this, please behave differently than we used
to" (because the status quo is "send a thin pack, whether I said it was
OK or not").

-Peff

Re: [PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Shawn Pearce]

On Wed, Nov 6, 2013 at 2:54 PM, Jeff King <peff@peff.net> wrote:
> If we instead introduced "no-thin", it is more like:
>
>   1. Receive-pack starts advertising "no-thin" (as dictated by
>      circumstances, as Carlos describes).
>
>   2. Send-pack which does not understand no-thin will ignore it and send
>      a thin pack. This is the same as now, and the same as step 2 above.
>
>   3. An upgraded send-pack will understand no-thin and do as the server
>      asks.
>
> So an upgraded client and server can start cooperating immediately, and
> we do not have to wait for the long assumption time to pass before
> applying the second half.
>
> It is tempting to think about a "thin" flag because that would be the
> natural way to have implemented it from the very beginning. But it is
> not the beginning, and the negative flag is the only way at this point
> to say "if you understand this, please behave differently than we used
> to" (because the status quo is "send a thin pack, whether I said it was
> OK or not").

I think the only sane option at this point is a "no-thin" flag, or
just require servers that want to be wire compatible to accept thin
packs.

Re: [PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Shawn Pearce]

On Wed, Nov 6, 2013 at 1:41 PM, Carlos Martín Nieto <cmn@elego.de> wrote:
> On Wed, 2013-11-06 at 12:32 -0800, Junio C Hamano wrote:
>> I'll queue these for now, but I doubt the wisdom of this series,
>> given that the ship has already sailed long time ago.
>>
>> Currently, no third-party implementation of a receiving end can
>> accept thin push, because "thin push" is not a capability that needs
>> to be checked by the current clients.  People will have to wait
>> until the clients with 2/2 patch are widely deployed before starting
>> to use such a receiving end that is incapable of "thin push".
>>
>> Wouldn't the world be a better place if instead they used that time
>> waiting to help such a third-party receiving end to implement "thin
>> push" support?
>>
>
> Support in the code isn't always enough. The particular case that
> brought this on is one where the index-pack implementation can deal with
> thin packs just fine.
>
> This particular service takes the pack which the client sent and does
> post-processing on it to store it elsewhere. During the receive-pack
> equivalent, there is no git object db that it can query for the missing
> base objects. I realise this is pretty a unusual situation.

How... odd?

At Google we have made effort to ensure servers can accept thin packs,
even though its clearly easier to accept non-thin, because clients in
the wild already send thin packs and changing the deployed clients is
harder than implementing the existing protocol.

If the server can't complete the pack, I guess this also means the
client cannot immediately fetch from the server it just pushed to?

Re: [PATCH 0/2] thin-pack capability for send-pack/receive-pack

[Carlos Martín Nieto]

On Wed, 2013-11-06 at 15:42 -0800, Shawn Pearce wrote:
> On Wed, Nov 6, 2013 at 1:41 PM, Carlos Martín Nieto <cmn@elego.de> wrote:
> > On Wed, 2013-11-06 at 12:32 -0800, Junio C Hamano wrote:
> >> I'll queue these for now, but I doubt the wisdom of this series,
> >> given that the ship has already sailed long time ago.
> >>
> >> Currently, no third-party implementation of a receiving end can
> >> accept thin push, because "thin push" is not a capability that needs
> >> to be checked by the current clients.  People will have to wait
> >> until the clients with 2/2 patch are widely deployed before starting
> >> to use such a receiving end that is incapable of "thin push".
> >>
> >> Wouldn't the world be a better place if instead they used that time
> >> waiting to help such a third-party receiving end to implement "thin
> >> push" support?
> >>
> >
> > Support in the code isn't always enough. The particular case that
> > brought this on is one where the index-pack implementation can deal with
> > thin packs just fine.
> >
> > This particular service takes the pack which the client sent and does
> > post-processing on it to store it elsewhere. During the receive-pack
> > equivalent, there is no git object db that it can query for the missing
> > base objects. I realise this is pretty a unusual situation.
> 
> How... odd?
> 
> At Google we have made effort to ensure servers can accept thin packs,
> even though its clearly easier to accept non-thin, because clients in
> the wild already send thin packs and changing the deployed clients is
> harder than implementing the existing protocol.

It is harder, but IMO also more correct, as thin packs are an
optimisation that was added somewhat later. Not to say it shouldn't be
something you should attempt to do, but it's a trade-off between the
complexity of the communication between the pieces and the potential
amount of extra data you're willing to put up with.

The Google (Code) servers don't just support thin packs, but for
upload-pack, they force it upon the client, which is quite frustrating
as it won't even tell you why it closes the connection but sends a 500
instead, but that's a different story.

> 
> If the server can't complete the pack, I guess this also means the
> client cannot immediately fetch from the server it just pushed to?

Not all the details have been worked out yet, but the new history should
be converted into the target format before reporting success and closing
the connection. The Git frontend/protocol is one way of putting data
into the system, but that's not its native data storage format. The
database where this is getting stored only has very limited knowledge of
git.

I'll reroll the series with "no-thin" as mentioned elsewhere in this
thread.

   cmn