On 01/26/2017 02:23 PM, Junio C Hamano wrote:
Jonathan Tan <jonathantanmy@xxxxxxxxxx> writes:Currently, while performing packfile negotiation [1], upload-pack allows clients to specify their desired objects only as SHA-1s. This causes: (a) vulnerability to failure when an object turns non-existent during negotiation, which may happen if, for example, upload-pack is provided by multiple Git servers in a load-balancing arrangement, and (b) dependence on the server first publishing a list of refs with associated objects. To eliminate (a) and take a step towards eliminating (b), teach upload-pack to support requests in the form of ref names and globs (in addition to the existing support for SHA-1s) through a new line of the form "want-ref <ref>" where ref is the full name of a ref, a glob pattern, or a SHA-1. At the conclusion of negotiation, the server will write "wanted-ref <SHA-1> <name>" for all requests that have been specified this way.I am not sure if this "at the conclusion of" is sensible. It is OK to assume that what the client side has is fixed, and it is probably OK to desire that what the server side has can change, but at the same time, it feels quite fragile to move the goalpost in between.
Do you have any specific concerns as to this fragility? Peff mentioned some concerns with the client making some decisions based on the initial SHA-1 vs the SHA-1 reported by "wanted-ref", to which I replied [1].
Stepping back a bit, in an environment that involves multiple server instances that have inconsistent set of refs, can the negotiation even be sensibly and safely implemented? The first server the client contacts may, in response to a "have", say "I do have that commit so you do not have to send its ancestors to me. We found one cut-off point. Please do explore other lines of histories." The next server that concludes the negotiation exchange may not have that commit and will be unable to produce a pack that excludes the objects reachable from that commit---wouldn't that become a problem?
It's true that this patch set wouldn't solve this problem. This problem only occurs when there is a commit that the client knows but only a few of the servers know (maybe because the client just pushed it to one of them). If, for example, the client does not know a commit and only a few of the servers know it (for example, because another user just pushed it), this patch set does help. The latter scenario seems like it would occur relatively commonly.
One way to prevent such a problem from hurting clients may be for these multiple server instances to coordinate and make sure they have a shared perception of the common history among them. Some pushes may have come to one instance but may not have propagated to other instances, and such a commit cannot be accepted as usable "have" if the servers anticipate that the final client request would go to any of the servers. Otherwise the multiple server arrangement would not work safely, methinks. And if the servers are ensuring the safety using such a mechanism, they can use the same mechanism to restrain "faster" instances from sending too fresh state of refs that other instances haven't caught up to, which would mean they can present a consistent set of refs to the client in the first place, no? So I am not sure if the mechanism to request history by refname instead of the tip commit would help the multi-server environment as advertised. It may help solving other problems, though (e.g. like "somebody pushed to update after the initial advertisement was sent out" which can happen even in a single server environment).
This patch set would solve the problem you describe (whether in a single server environment or the coordination between multiple servers that provides "strong consistency"). (Although it may not be an important problem to solve, since it is probably OK if the client got a "slow" version of the state of the refs.)
To be flexible with respect to client needs, the server does not indicate an error if a "want-ref" line corresponds to no refs, but instead relies on the client to ensure that what the user needs has been fetched. For example, a client could reasonably expand an abbreviated name "foo" to "want-ref foo", "want-ref refs/heads/foo", "want-ref refs/tags/foo", among others, and ensure that at least one such ref has been fetched.Cute. This may be one way to implement the DWIM thing within the constraint of eventually wanting to go to "client speaks first, the server does not advertise things the client is not interested in" world. But at the same time it may end up bloating the set of refs the client asks instead. Instead of receiving the advertisement and then sending one request after picking the matching one from it, the client needs to send "refs/{heads,tags,whatever}/foo".
That is true, although I think that the client will typically send only a few ref names (with or without globs), so the request packet is still not that large.
[1] <67afbb3b-5d0b-8c0d-3f6e-3f559c68f4bd@xxxxxxxxxx>