Merge branch '18874-merge-wb2'
[arvados.git] / lib / dispatchcloud / worker / worker.go
index 8b9326fa9c0025467fc85578cf3b0e614dbe4a55..10a28157e43ee0e496fd82b8705542a58aa0c8ca 100644 (file)
@@ -188,6 +188,14 @@ func (wkr *worker) startContainer(ctr arvados.Container) {
                }
                wkr.mtx.Lock()
                defer wkr.mtx.Unlock()
+               if wkr.starting[ctr.UUID] != rr {
+                       // Someone else (e.g., wkr.probeAndUpdate() ->
+                       // wkr.updateRunning() or wkr.Close()) already
+                       // moved our runner from wkr.starting to
+                       // wkr.running or deleted it while we were in
+                       // rr.Start().
+                       return
+               }
                now := time.Now()
                wkr.updated = now
                wkr.busy = now
@@ -547,9 +555,11 @@ func (wkr *worker) shutdownIfBroken(dur time.Duration) bool {
                // Never shut down.
                return false
        }
-       label, threshold := "", wkr.wp.timeoutProbe
+       prologue, epilogue, threshold := "", "", wkr.wp.timeoutProbe
        if wkr.state == StateUnknown || wkr.state == StateBooting {
-               label, threshold = "new ", wkr.wp.timeoutBooting
+               prologue = "new "
+               epilogue = " -- `arvados-server cloudtest` might help troubleshoot, see https://doc.arvados.org/main/admin/cloudtest.html"
+               threshold = wkr.wp.timeoutBooting
        }
        if dur < threshold {
                return false
@@ -558,7 +568,7 @@ func (wkr *worker) shutdownIfBroken(dur time.Duration) bool {
                "Duration": dur,
                "Since":    wkr.probed,
                "State":    wkr.state,
-       }).Warnf("%sinstance unresponsive, shutting down", label)
+       }).Warnf("%sinstance unresponsive, shutting down%s", prologue, epilogue)
        wkr.shutdown()
        return true
 }
@@ -665,10 +675,12 @@ func (wkr *worker) Close() {
        for uuid, rr := range wkr.running {
                wkr.logger.WithField("ContainerUUID", uuid).Info("crunch-run process abandoned")
                rr.Close()
+               delete(wkr.running, uuid)
        }
        for uuid, rr := range wkr.starting {
                wkr.logger.WithField("ContainerUUID", uuid).Info("crunch-run process abandoned")
                rr.Close()
+               delete(wkr.starting, uuid)
        }
 }