On 8/5/2021 11:07 AM, Ævar Arnfjörð Bjarmason wrote:> We're in the 2.33.0 rc cycle, and I'd hoped to have some more prep > work for this integrated already, but for now here's something > interesting I've been working on for early commentary/feedback. I've taking a brief look at the code, but mostly have thoughts about the core idea based on reading the documentation. I'll keep my feedback here restricted to that, especially because I saw some thoughts from Jonathan that question the idea. > This adds the the ability to protocol v2 for servers to optimistically > pre-seed supporting clients with one or more bundles via a new > "bundle-uri" protocol extension. To make sure I understand things correctly, I'm going to rewrite the description of the feature in my own words. Please correct me if I have misread something. This presents a new capability for protocol v2 that allows a server to notify a client that they can download a bundle to bootstrap their object database, and then come back to the origin server to "catch up" the remaining new objects since that bundle via a fetch negotiation. This idea is similar to the packfile-uri feature in that it offloads some server load to a dumb data store (like a CDN) but differs in a few key ways: 1. The bundle download does not need to be controlled by the server, and the server doesn't even need to know its contents. The client will discover the content and share its view in a later negotiation. 2. The packfile-uri could advertise a pack that contains only the large reachable blobs, and hence could use the "happy path" for bitmaps to compute a packfile containing all reachable objects except these large blobs. For the bundle-uri feature to mimic this, the blobs would need to be reachable from refs (perhaps shallow tags?) but it would not be clear that the negotiation would prevent redownloading those files. This is an area to explore and expand upon. 3. The bundle-uri feature is focused on "clone" scenarios, and I don't see how it could be used to help "fetch" scenarios. To be fair, I also have been vocal about how the packfile-uri feature is very hard to make helpful for the "fetch" scenario. The benefits to "clone" seem to be worth the effort alone. I think the bundle-api doubles-down on that being the core functionality that matters, and that is fine by me. It sacrifices the flexibility of the packfile-uri with a lower maintenance burden for servers that want to implement it. The biggest benefit that I see is that the Git server does not need to worry about whether the bundle has an exact set of data that it expects. There is no timing issue about whether or not the exact packfile is seeded. Instead, the server could have a fixed URL for its bundle and update its contents (atomically) at any schedule without worrying about concurrent clones being interrupted. From an operational perspective, I find the bundle-uri a better way to offload "clone" load. This also depends on that following "git fetch" being easy to serve. In that sense, it can be beneficial to be somewhat aware of the bundles that are being served: can we store the bundled refs as reachability bitmaps so we have those available for the negotiation in the following "git fetch" operations? This choice seems specific to how the server is deciding to create these bundles. It also presents interesting ideas for how to create the bundle to focus on some core portion of the repo. The "thundering herd" of CI machines that re-clone repos at a high rate will also be likely to use the "--single-branch" option to reduce the refs that they will ask for in the negotiation. In that sense, we won't want a snapshot of all refs at a given time and instead might prefer a snapshot of the default branch or some set of highly-active branches. One question I saw Jonathan ask was "can we modify the packfile-uri capability to be better?" I think this feature has enough different goals that they could co-exist. That's the point of protocol v2, right? Servers can implement and advertise the subset of functionality that they think is best for their needs. I hope my ramblings provide some amount of clarity to the discussion, but also I intend to show support of the idea. If I was given the choice of which feature to support (I mostly work on the client experience, so take my choice with a grain of salt), then I would focus on implementing the bundle-uri capability _before_ the packfile-uri capability. And that's the best part: more options present more flexibility for different hosts to make different decisions. Thanks, -Stolee
