Fix 5s delay on sticky queue (#2811)

Sticky worker mark workflows as running on sticky queue. But when sticky worker
is gone (gracefully shutdown or forcefully killed), they don't notify server. 
When new workflow task or query is scheduled, server still dispatch them to the 
old sticky queue which now has no worker working on and will only result in 5s 
sticky queue's schedule_to_start timeout. No mater how long after the old 
sticky worker is gone, there is always that 5s delay. This PR fix that by 
checking if if there is no inflight poller for the sticky queue, and no poller 
is seem within 10s. In that case, matching would return a new 
StickyWorkerUnavailable error and history would retry that by using original 
task queue.

client/history/metricClient.go

@@ -632,6 +632,7 @@ func (c *metricClient) finishMetricsRecording(
 		case *serviceerror.Canceled,
 			*serviceerror.DeadlineExceeded,
 			*serviceerror.NotFound,
+			*serviceerror.QueryFailed,
 			*serviceerror.NamespaceNotFound,
 			*serviceerror.WorkflowExecutionAlreadyStarted:
 			// noop - not interest and too many logs

client/matching/metricClient.go

@@ -30,6 +30,7 @@ import (
 
 	"go.temporal.io/api/serviceerror"
 	taskqueuepb "go.temporal.io/api/taskqueue/v1"
+	serviceerrors "go.temporal.io/server/common/serviceerror"
 	"google.golang.org/grpc"
 
 	"go.temporal.io/server/api/matchingservice/v1"
@@ -258,10 +259,12 @@ func (c *metricClient) finishMetricsRecording(
 ) {
 	if err != nil {
 		switch err.(type) {
-		case *serviceerror.Canceled,
+		case *serviceerrors.StickyWorkerUnavailable,
+			*serviceerror.Canceled,
 			*serviceerror.DeadlineExceeded,
 			*serviceerror.NotFound,
 			*serviceerror.NamespaceNotFound,
+			*serviceerror.QueryFailed,
 			*serviceerror.WorkflowExecutionAlreadyStarted:
 			// noop - not interest and too many logs
 		default:

common/rpc/interceptor/telemetry.go

@@ -159,6 +159,9 @@ func (ti *TelemetryInterceptor) handleError(
 	}
 
 	switch err := err.(type) {
+	case *serviceerrors.StickyWorkerUnavailable:
+		// we emit service_errors_with_type metrics, no need to emit specific metric for this error type.
+		// TODO deprecate all metrics below
 	case *serviceerrors.ShardOwnershipLost:
 		scope.IncCounter(metrics.ServiceErrShardOwnershipLostCounter)
 	case *serviceerrors.TaskAlreadyStarted:

common/serviceerror/convert.go

@@ -58,6 +58,11 @@ func FromStatus(st *status.Status) error {
 		case *errordetails.RetryReplicationFailure:
 			return newRetryReplication(st, errDetails)
 		}
+	case codes.Unavailable:
+		switch errDetails.(type) {
+		case *errordetails.StickyWorkerUnavailableFailure:
+			return newStickyWorkerUnavailable(st)
+		}
 	}
 
 	return serviceerror.FromStatus(st)