On Sun, Nov 04, 2018 at 03:50:17AM +0000, Eric Wong wrote: > > Maybe, but If "git clone" stresses the server to the point of breakage > > as you say, something should be done on the server side. I've been > > using the kernel.org repo as Konstantin suggested. > > Based on Konstantin's message; it didn't seem like the server > itself was stressed, just something on Amazon's side.
Right, it's a bit of an unfixable stupidity on the AWS side of things that manifests itself when cloning very large repositories (search for "git clone aws connection draining" -- and keep in mind that we've tried all of those workarounds). The only fix for this is to switch from using AWS ALB (http-level proxy) to an NLB (tcp-level proxy), but then we lose some niceties like certificate management. We'll probably just end up setting cloneurl to direct people to git.kernel.org clone locations instead. > > Perhaps lower the recommended maximum shard size? I didn't hit > > this issue with the most recent shard, which is about 50% smaller. > > That's another option; but migrating/converting existing repos > to it would be a flag day and break existing mirrors. > ~1G seemed like a reasonable starting point It's not really the size, but how long the client takes fetching the repository objects, which is why the problem only manifests itself to a subset of people. Best, -K