Found new dashboard test flakes for:

#!watchflakes
default <- builder == "x_tools-go1.22-darwin-amd64-longtest" && repo == "tools" && mode == "build"

ERROR: rpc error: code = NotFound desc = stream path not found

— watchflakes

Looks like a LUCI infrastructure problem?
(attn @mknyszek @dmitshur)

I think watchflakes is reporting this issue incorrectly, it should be mentioning the "signal: killed" error instead.

It's not really clear to me what went wrong here. It looks like the command failed immediately (8 ms of execution time before failure) which suggests that it's not go test failing but a tool that's part of the infrastructure (possibly rdb or result_adapter, though I suppose it could get to go test in 8 ms). I've been meaning to break out the tool invocations from the test invocation, which would allow us to mark this correctly as an infra failure, but it's annoying to do so because it's all structured as wrapper commands.

Anyway, it seems the subprocess got so forcefully killed that it's hard to say what went wrong. Out of memory perhaps? Or a Darwin bug?

Lastly, I'll point out that this ran on a canary version of LUCI, which might be interesting, but I think a subprocess getting killed like this seems unlikely to occur from that. Specifically, we fetch the tools we wrap tests with ourselves.