Shutdown state server and metric manager before service layer

Avoids a very small race condition window where metric update hooks
may point into the service layer components even after they have
been destroyed, as these are not explicitly unregistered today.

Another option would be to add unregistering on component destruction,
but that adds another race condition where an external client may
observe partial metric existence during this time window. By shutting
down the metric exporting interfaces first, we should avoid this.

1 files changed, 3 insertions, 0 deletions

diff --git a/metrics/src/vespa/metrics/metricmanager.cpp b/metrics/src/vespa/metrics/metricmanager.cpp
index 7ac912416de..84e92c2c316 100644
--- a/metrics/src/vespa/metrics/metricmanager.cpp
+++ b/metrics/src/vespa/metrics/metricmanager.cpp
@@ -94,6 +94,9 @@ MetricManager::~MetricManager()
 void
 MetricManager::stop()
 {
+    if (!running()) {
+        return; // Let stop() be idempotent.
+    }
     Runnable::stop();
     {
         vespalib::MonitorGuard sync(_waiter);