exchangev2: fetch file revisions Now that the server has an API for fetching file data, we can call into it to fetch file revisions. The implementation is relatively straightforward: we examine the manifests that we fetched and find all new file revisions referenced by them. We build up a mapping from file path to file nodes to manifest node. (The mapping to first manifest node allows us to map back to first changelog node/revision, which is used for the linkrev.) Once that map is built up, we iterate over it in a deterministic manner and fetch and store file data. The code is very similar to manifest fetching. So similar that we could probably extract the common bits into a generic function. With file data retrieval implemented, `hg clone` and `hg pull` are effectively feature complete, at least as far as the completeness of data transfer for essential repository data (changesets, manifests, files, phases, and bookmarks). We're still missing support for obsolescence markers, the hgtags fnodes cache, and the branchmap cache. But these are non-essential for the moment (and will be implemented later). This is a good point to assess the state of exchangev2 in terms of performance. I ran a local `hg clone` for the mozilla-unified repository using both version 1 and version 2 of the wire protocols and exchange methods. This is effectively comparing the performance of the wire protocol overhead and "getbundle" versus domain-specific commands. Wire protocol version 2 doesn't have compression implemented yet. So I tested version 1 with `server.compressionengines=none` to remove compression overhead from the equation. server before: user 220.420+0.000 sys 14.420+0.000 after: user 321.980+0.000 sys 18.990+0.000 client before: real 561.650 secs (user 497.670+0.000 sys 28.160+0.000) after: real 1226.260 secs (user 944.240+0.000 sys 354.150+0.000) We have substantial regressions on both client and server. This is obviously not desirable. I'm aware of some reasons: * Lack of hgtagsfnodes transfer (contributes significant CPU to client). * Lack of branch cache transfer (contributes significant CPU to client). * Little to no profiling / optimization performed on wire protocol version 2 code. * There appears to be a memory leak on the client and that is likely causing swapping on my machine. * Using multiple threads on the client may be counter-productive because Python. * We're not compressing on the server. * We're tracking file nodes on the client via manifest diffing rather than using linkrev shortcuts on the server. I'm pretty confident that most of these issues are addressable. But even if we can't get wire protocol version 2 on performance parity with "getbundle," I still think it is important to have the set of low level data-specific retrieval commands that we have implemented so far. This is because the existence of such commands allows flexibility in how clients access server data. Differential Revision: https://phab.mercurial-scm.org/D4491

wireprotov2: define and implement "filedata" command Continuing our trend of implementing *data commands for retrieving information about specific repository data primitives, this commit implements a command for retrieving data about an individual tracked file. The command is very similar to "manifestdata." The only significant difference is that we have a standalone function for obtaining storage for a tracked file. This is to provide a monkeypatch point for extensions to implement path-based access control. With this API available, wire protocol version 2 now exposes all data primitives necessary to implement a full clone. Of course, since "filedata" can only resolve data for a single path at a time, clients would need to issue N commands to perform a full clone. On the Firefox repository, this would be ~461k commands. We'll likely need to implement a file data retrieval command that supports multiple paths. But that can be implemented later. Differential Revision: https://phab.mercurial-scm.org/D4490

exchangev2: fetch manifest revisions Now that the server has support for retrieving manifest data, we can implement the client bits to call it. We teach the changeset fetching code to capture the manifest revisions that are encountered on incoming changesets. We then feed this into a new function which filters out known manifests and then batches up manifest data requests to the server. This is different from the previous wire protocol in a few notable ways. First, the client fetches manifest data separately and explicitly. Before, we'd ask the server for data pertaining to some changesets (via a "getbundle" command) and manifests (and files) would be sent automatically. Providing an API for looking up just manifest data separately gives clients much more flexibility for manifest management. For example, a client may choose to only fetch manifest data on demand instead of prefetching it (i.e. partial clone). Second, we send N commands to the server for manifest retrieval instead of 1. This property has a few nice side-effects. One is that the deterministic nature of the requests lends itself to server-side caching. For example, say the remote has 50,000 manifests. If the server is configured to cache responses, each time a new commit arrives, you will have a cache miss and need to regenerate all outgoing data. But if you makes N requests requesting 10,000 manifests each, a new commit will still yield cache hits on the initial, unchanged manifest batches/requests. A derived benefit from these properties is that resumable clone is conceptually simpler to implement. When making a monolithic request for all of the repository data, recovering from an interrupted clone is hard because the server was in the driver's seat and was maintaining state about all the data that needed transferred. With the client driving fetching, the client can persist the set of unfetched entities and retry/resume a fetch if something goes wrong. Or we can fetch all data N changesets at a time and slowly build up a repository. This approach is drastically easier to implement when we have server APIs exposing low-level repository primitives (such as manifests and files). We don't yet support tree manifests. But it should be possible to implement that with the existing wire protocol command. Differential Revision: https://phab.mercurial-scm.org/D4489

wireprotov2: define and implement "manifestdata" command The added command can be used for obtaining manifest data. Given a manifest path and set of manifest nodes, data about manifests can be retrieved. Unlike changeset data, we wish to emit deltas to describe manifest revisions. So the command uses the relatively new API for building delta requests and emitting them. The code calls into deltaparent(), which I'm not very keen of. There's still work to be done in delta generation land so implementation details of storage (e.g. exactly one delta is stored/available) don't creep into higher levels. But we can worry about this later (there is already a TODO on imanifestorage tracking this). On the subject of parent deltas, the server assumes parent revisions exist on the receiving end. This is obviously wrong for shallow clone. I've added TODOs to add a mechanism to the command to allow clients to specify desired behavior. This shouldn't be too difficult to implement. Another big change is that the client must explicitly request manifest nodes to retrieve. This is a major departure from "getbundle," where the server derives relevant manifests as it iterates changesets and sends them automatically. As implemented, the client must transmit each requested node to the server. At 20 bytes per node, we're looking at 2 MB per 100,000 nodes. Plus wire encoding overhead. This isn't ideal for clients with limited upload bandwidth. I plan to address this in the future by allowing alternate mechanisms for defining the revisions to retrieve. One idea is to define a range of changeset revisions whose manifest revisions to retrieve (similar to how "changesetdata" works). We almost certainly want an API to look up an individual manifest by node. And that's where I've chosen to start with the implementation. Again, a theme of this early exchangev2 work is I want to start by building primitives for accessing raw repository data first and see how far we can get with those before we need more complexity. Differential Revision: https://phab.mercurial-scm.org/D4488

wireprotov2: add TODOs around extending changesetdata fields Extensions will inevitably want to extend the set of changeset data/fields that can be requested. We'll need to implement support for extending this in the future. Add some TODOs to track that. Differential Revision: https://phab.mercurial-scm.org/D4487

exchangev2: fetch and apply bookmarks This is pretty similar to phases data. We collect bookmarks data as we process records. Then at the end we make a call to the bookmarks subsystem to reflect the remote's bookmarks. Like phases, the code for handling bookmarks is vastly simpler than the previous wire protocol code because the server always transfers the full set of bookmarks when bookmarks are requested. We don't have to keep track of whether we requested bookmarks or not. Differential Revision: https://phab.mercurial-scm.org/D4486

wireprotov2: add bookmarks to "changesetdata" command Like we did for phases, we want to emit bookmarks data attached to each changeset. The approach here is very similar to phases: we emit bookmarks data inline with requested revision data. But we emit records for nodes that weren't requested as well so consumers have access to the full set of defined bookmarks. Differential Revision: https://phab.mercurial-scm.org/D4485

exchangev2: fetch and apply phases data Now that the server supports emitting phases data, we can request it and apply it on the client. Because we may receive phases-only updates from the server, we no longer conditionally perform the "changesetdata" command depending on whether there are revisions to fetch. In the previous wire protocol, this case would result in us falling back to performing "listkeys" commands to look up phases, bookmarks, etc data. But since "changesetdata" is smart enough to handle metadata only fetches, we can keep things consistent. It's worth noting that because of the unified approach to changeset data retrieval, phase handling code in wire proto v2 exchange is drastically simpler. Contrast with all the code in exchange.py dealing with all the variations for obtaining phases data. Differential Revision: https://phab.mercurial-scm.org/D4484

wireprotov2: add phases to "changesetdata" command This commit teaches the "changesetdata" wire protocol command to emit the phase state for each changeset. This is a different approach from existing phase transfer in a few ways. Previously, if there are no new revisions (or we're not using bundle2), we perform a "listkeys" request to retrieve phase heads. And when revision data is being transferred with bundle2, phases data is encoded in a standalone bundle2 part. In both cases, phases data is logically decoupled from the changeset data and is encountered/applied after changeset revision data is received. The new wire protocol purposefully tries to more tightly associate changeset metadata (phases, bookmarks, obsolescence markers, etc) with the changeset revision and index data itself, rather than have it live as a separate entity that must be fetched and processed separately. I reckon that one reason we didn't do this before was it was difficult to add new data types/fields without breaking existing consumers. By using CBOR maps to transfer changeset data and putting clients in control of what fields are requested / present in those maps, we can easily add additional changeset data while maintaining backwards compatibility. I believe this to be a superior approach to the problem. That being said, for performance reasons, we may need to resort to alternative mechanisms for transferring data like phases. But for now, I think giving the wire protocol the ability to transfer changeset metadata next to the changeset itself is a powerful feature because it is a raw, changeset-centric data API. And if you build simple APIs for accessing the fundamental units of repository data, you enable client-side experimentation (partial clone, etc). If it turns out that we need specialized APIs or mechanisms for transferring data like phases, we can build in those APIs later. For now, I'd like to see how far we can get on simple APIs. It's worth noting that when phase data is being requested, the server will also emit changeset records for nodes in the bases specified by the "noderange" argument. This is to ensure that phase-only updates for nodes the client has are available to the client, even if no new changesets will be transferred. Differential Revision: https://phab.mercurial-scm.org/D4483

exchangev2: fetch changeset revisions All Mercurial repository data is derived from changesets: you can't do anything unless you have changesets. Therefore, it makes sense for changesets to be the first piece of data that we transfer as part of pull. To do this, we call our new "changesetdata" command, requesting parents and revision data. This gives us all the data that a changegroup delta group would give us. We simply normalize this data into what addgroup() expects and call that API on the changelog to bulk insert revisions into the changelog. Code in this commit is heavily borrowed from changegroup.cg1unpacker.apply(). Differential Revision: https://phab.mercurial-scm.org/D4482