A number of clusters in the wild on 4.1.z (15-20?) are reporting one etcd member down via the `up` metric, but no alerts related to etcd failure are being reported. Other clusters with one etcd member reported down ARE reporting alerts related to a bad member:
80f5da7e-7527-41d2-8d6e-774b388a42a4 reports the following alerts:
KubeDeploymentReplicasMismatch, KubePodNotReady, TargetDown, Watchdog
and two down services
This is a UPI cluster at 4.1.4. We may have a scheduling issue with the etcd proxy, but more data needs to be gathered.
Ok, looking into this we have a gap in our alerts.
Today we have etcdInsufficientMembers which depends on the etcd service being present but down (we must have 3 up series). However, when this happens we only fire TargetDown and that's concerning since we're really degraded. We need to have a better alert for degradation
Second, there is a different failure mode where a node is completely removed, at which point the number of series in up is 2, and the alerts do not fire correctly. This is a valid failure mode that our alerts don't cover. The query `count(sum by (To) (rate(etcd_network_peer_sent_failures_total[2m])) > 0) > 0` roughly approximates this.
Verfied with 4.1.0-0.nightly-2019-08-06-212225,
shutdown 1 etcd member, then check Monitoring section in web console, got expected alert msg:
etcd cluster "etcd": members are down (1).
2 minutes ago
Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.
For information on the advisory, and where to find the updated
files, follow the link below.
If the solution does not work for you, open a new bug report.