Skipping CI for Draft Pull Request.
If you want CI signal for your change, please convert it to an actual PR.
You can still manually trigger a test run with /test all

@dcantah I think you've played with grpc stuff recently, pls PTAL.

@dmcgowan also as we were chatting on the history of a lot of these options on slack a bit ago. We were wondering if we would be able to just forego WithBlock today like we do here

I believe the original issue like 8 years was related to clients connecting and when the daemon was not online the client would end up hanging. Now in that case it seems to be handled better with this PR.

In this PR...

$ time bin/ctr images ls
WARN[0001] Failed to check deprecations                  error="connection error: desc = \"transport: Error while dialing: dial unix:///run/containerd/containerd.sock: timeout\": unavailable"
ctr: failed to list images: connection error: desc = "transport: Error while dialing: dial unix:///run/containerd/containerd.sock: timeout": unavailable
bin/ctr images ls  0.02s user 0.03s system 4% cpu 1.019 total

it comes back in 1 second. Although would be good to avoid the warning log.

In main...

time bin/ctr images ls
ctr: failed to dial "/run/containerd/containerd.sock": context deadline exceeded: connection error: desc = "transport: error while dialing: dial unix:///run/containerd/containerd.sock: timeout"
bin/ctr images ls  0.07s user 0.08s system 1% cpu 10.019 total

it takes 10 seconds to timeout.

Windows needed workaround 20aebe7. Everything else looks green.

Although would be good to avoid the warning log.

@dmcgowan The warning is coming from an extra call to introspection service when creating a client in ctr, which we always do. It's weird that you don't see it the second case (unless it was suppressed or succeeded?).

@mxpv @dcantah Is the net.DialTimeout that was introduced by #9447 still needed?

@ZhangShuaiyi Good question, I'm not sure. The reason for it was so that on restart we can tell what shims to ignore if they're not responding when going to reload state (usually because they crashed for some reason so the socket isn't listening). We already had that info with grpc.WithBlock, it would just take the entire timeout to figure out if we should ignore the shim so we added that to be able to find out much faster if the socket even exists anymore. Now with this scheme I'm not sure if we know whether we can ignore a shim or not when reloading, as we don't do any network ops when making a client anymore, so maybe that dial still has value?

PR needs rebase.

@kzys Could you enqueue this again?