From: Tom Clegg <tclegg@veritasgenetics.com>
Date: Mon, 17 Sep 2018 18:33:59 +0000 (-0400)
Subject: 13994: Fix service discovery race.
X-Git-Tag: 1.3.0~108^2~2
X-Git-Url: https://git.arvados.org/arvados.git/commitdiff_plain/ddc180fa5200d9d8fac59cc5041d7d452b68e6a2

13994: Fix service discovery race.

Making keepclient.RefreshServiceDiscovery asynchronous didn't work
out: tests rely on it to ensure old services won't be used.

Instead, speed things up a bit by refreshing all known clients'
service lists concurrently instead of serially.

Arvados-DCO-1.1-Signed-off-by: Tom Clegg <tclegg@veritasgenetics.com>
---

diff --git a/sdk/go/keepclient/discover.go b/sdk/go/keepclient/discover.go
index 6778b39fbc..4377c19515 100644
--- a/sdk/go/keepclient/discover.go
+++ b/sdk/go/keepclient/discover.go
@@ -22,12 +22,15 @@ import (
 func RefreshServiceDiscovery() {
 	svcListCacheMtx.Lock()
 	defer svcListCacheMtx.Unlock()
+	var wg sync.WaitGroup
 	for _, ent := range svcListCache {
-		select {
-		case ent.clear <- struct{}{}:
-		default:
-		}
+		wg.Add(1)
+		go func() {
+			ent.clear <- struct{}{}
+			wg.Done()
+		}()
 	}
+	wg.Wait()
 }
 
 // ClearCacheOnSIGHUP installs a signal handler that calls
@@ -139,7 +142,7 @@ func (kc *KeepClient) discoverServices() error {
 		arv := *kc.Arvados
 		cacheEnt = cachedSvcList{
 			latest: make(chan svcList),
-			clear:  make(chan struct{}, 1),
+			clear:  make(chan struct{}),
 			arv:    &arv,
 		}
 		go cacheEnt.poll()