From c6bc053a676ee8dceefeb072d0b17cc5f5df03c4 Mon Sep 17 00:00:00 2001
From: irfan sharif <irfanmahmoudsharif@gmail.com>
Date: Mon, 31 Aug 2020 17:58:51 -0400
Subject: [PATCH] server: busy loop through resolver list during join process

Deferred doing this in #52526. Probably a good idea to do have it, it'll
bring down the cluster convergence time (time taken for all nodes to
find out about the initialization) by a bit.

Release justification: low risk, high benefit changes to existing functionality
Release note: None
---
 pkg/server/init.go | 33 +++++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/pkg/server/init.go b/pkg/server/init.go
index d2932c8f595e..f915b81f7657 100644
--- a/pkg/server/init.go
+++ b/pkg/server/init.go
@@ -406,6 +406,39 @@ func (s *initServer) startJoinLoop(ctx context.Context, stopper *stop.Stopper) (
 		return nil, errJoinRPCUnsupported
 	}
 
+	// Iterate through all the resolvers at least once to reduce time taken to
+	// cluster convergence. Keep this code block roughly in sync with the one
+	// below.
+	for _, res := range s.config.resolvers {
+		select {
+		case <-ctx.Done():
+			return nil, context.Canceled
+		case <-stopper.ShouldQuiesce():
+			return nil, stop.ErrUnavailable
+		default:
+		}
+
+		addr := res.Addr()
+		state, err := s.attemptJoinTo(ctx, res.Addr())
+		if err == nil {
+			return state, nil
+		}
+
+		if errors.Is(err, errJoinRPCUnsupported) || errors.Is(err, ErrIncompatibleBinaryVersion) {
+			// Propagate upwards; these are error conditions the caller knows to
+			// expect.
+			return nil, err
+		}
+
+		if IsWaitingForInit(err) {
+			log.Warningf(ctx, "%s is itself waiting for init, will retry", addr)
+		} else {
+			log.Warningf(ctx, "outgoing join rpc to %s unsuccessful: %v", addr, err.Error())
+		}
+
+		// Try the next node if unsuccessful.
+	}
+
 	const joinRPCBackoff = time.Second
 	var tickChan <-chan time.Time
 	{