Mercurial: Changelog

wireprotov2: server support for sending content redirects A "content redirect" can be sent in place of inline response content. In terms of code, we model a content redirect as a special type of response object holding the attributes describing that redirect. Sending a content redirect thus becomes as simple as the object emission layer sending an instance of that type. A cacher using externally-addressable content storage could replace the outgoing object stream with an object advertising its location. The bulk of the code in this commit is teaching the output layer which handles the object stream to recognize alternate location objects. The rules are that if an alternate location object is present, it must be the first and only object in the object stream. Otherwise the server emits an error. Differential Revision: https://phab.mercurial-scm.org/D4777

wireprotov2: client support for advertising redirect targets With the server now able to emit a redirect target descriptor, we can start to teach the client to recognize it. This commit implements support for filtering the advertised redirect targets against supported features and for advertising compatible redirect targets as part of command requests. It also adds the minimal boilerplate required to fail when a content redirect is seen. The server doesn't yet do anything with the advertised redirect targets. And the client can't yet follow redirects if it did. But at least we're putting bytes on the wire. Differential Revision: https://phab.mercurial-scm.org/D4776

wireprotov2: advertise redirect targets in capabilities This is pretty straightforward. Redirect targets will require an extension to support. So we've added a function that can be wrapped to define redirect targets. To test this, we teach our simple cache test extension to read redirect targets from a file. It's a bit hacky. But it gets the job done. Differential Revision: https://phab.mercurial-scm.org/D4775

wireprotov2: define semantics for content redirects When I implemented the clonebundles feature and deployed it on hg.mozilla.org using Amazon S3 as a content server, server-side CPU and bandwidth usage dropped off a cliff and a ton of server scaling headaches went away pretty much the instant clients with support for clonebundles were rolled out to Firefox CI. An obvious takeaway from that experience was that offloading server load to scalable file servers - potentially backed by a CDN - is a really good idea. Another takeaway was that Mercurial's wire protocol wasn't in a good position to support data offload generally. In wire protocol version 1, there isn't a mechanism in the protocol to say "grab the data from over here instead." For HTTP, we could teach the client to follow HTTP redirects. Or we could invent a media type that encoded redirects inline. But for SSH, we were pretty much out of luck because that protocol wasn't very flexible. Wire protocol version 2 offers the opportunity to do something better. The recent generic server-side content caching layer in the wire protocol version 2 server demonstrated that it is possible to have drop-in caching of responses to command requests. This by itself adds tons of value and already makes the built-in server much more scalable. But I don't want to stop there. The existing server-side caching implementation has a big weakness: it requires the server to send data to the client. This means that the Mercurial server is potentially sending gigabytes of data to thousands of clients. This is problematic because compared to scaling static file servers, scaling dynamic servers is *hard*. A solution to this is to "offload" serving of content to something that isn't the Mercurial server. By offloading content serving, you turn the Mercurial server from a centralized monolithic service to a distributed mostly-indexing service. Assuming high rates of content offload, this should drastically reduce the total work performed by the Mercurial server, both in terms of CPU and data transfer. This will make Mercurial servers vastly easier to scale. This commit defines the semantics for "content redirects" in wire protocol version 2. Essentially: * Servers advertise the set of locations a response could be served from. * When making requests, clients advertise the set of locations they are willing to fetch content from. * Servers can then replace the inline response with one that says "get the response from over here instead." This feature - when fully implemented - will allow extending the server-side caching layer to facilitate such things as integrating your server-side cache with a scalable blob store (such as S3 or a CDN) and offloading most data transfer to that external service. This feature could also be leveraged for load balancing. e.g. requests could come into a central server and then get redirected to an available mirror depending on server availability or locality. There's tons of potential :) Differential Revision: https://phab.mercurial-scm.org/D4774

wireprotov2: support response caching One of the things I've learned from managing VCS servers over the years is that they are hard to scale. It is well known that some companies have very beefy (read: very expensive) servers to power their VCS needs. It is also known that specialized servers for various VCS exist in order to facilitate scaling servers. (Mercurial is in this boat.) One of the aspects that make a VCS server hard to scale is the high CPU load incurred by constant client clone/pull operations. To alleviate the scaling pain associated with data retrieval operations, I want to integrate caching into the Mercurial wire protocol server as robustly as possible such that servers can aggressively cache responses and defer as much server load as possible. This commit represents the initial implementation of a general caching layer in wire protocol version 2. We define a new interface and behavior for a wire protocol cacher in repository.py. (This is probably where a reviewer should look first to understand what is going on.) The bulk of the added code is in wireprotov2server.py, where we define how a command can opt in to being cached and integrate caching into command dispatching. From a very high-level: * A command can declare itself as cacheable by providing a callable that can be used to derive a cache key. * At dispatch time, if a command is cacheable, we attempt to construct a cacher and use it for serving the request and/or caching the request. * The dispatch layer handles the bulk of the business logic for caching, making cachers mostly "dumb content stores." * The mechanism for invalidating cached entries (one of the harder parts about caching in general) is by varying the cache key when state changes. As such, cachers don't need to be concerned with cache invalidation. Initially, we've hooked up support for caching "manifestdata" and "filedata" commands. These are the simplest to cache, as they should be immutable over time. Caching of commands related to changeset data is a bit harder (because cache validation is impacted by changes to bookmarks, phases, etc). This will be implemented later. (Strictly speaking, censoring a file should invalidate caches. I've added an inline TODO to track this edge case.) To prove it works, this commit implements a test-only extension providing in-memory caching backed by an lrucachedict. A new test showing this extension behaving properly is added. FWIW, the cacher is ~50 lines of code, demonstrating the relative ease with which a cache can be added to a server. While the test cacher is not suitable for production workloads, just for kicks I performed a clone of just the changeset and manifest data for the mozilla-unified repository. With a fully warmed cache (of just the manifest data since changeset data is not cached), server-side CPU usage dropped from ~73s to ~28s. That's pretty significant and demonstrates the potential that response caching has on server scalability! Differential Revision: https://phab.mercurial-scm.org/D4773