Etcd Grpcrequests Slow

etcdGRPCRequestsSlow #

Meaning #

This alert fires when the 99th percentile of etcd gRPC requests are too slow.

Impact #

When requests are too slow, they can lead to various scenarios like leader election failure, slow reads and writes.

Diagnosis #

This could be result of slow disk (due to fragmented state) or CPU contention.

Slow disk #

One of the most common reasons for slow gRPC requests is disk. Checking disk related metrics and dashboards should provide a more clear picture.

PromQL queries used to troubleshoot #

Verify the value of how slow the etcd gRPC requests are by using the following query in the metrics console:

histogram_quantile(0.99, sum(rate(grpc_server_handling_seconds_bucket{job=~".*etcd.*", grpc_type="unary"}[5m])) without(grpc_type))

That result should give a rough timeline of when the issue started.

etcd_disk_wal_fsync_duration_seconds_bucket reports the etcd disk fsync duration, etcd_server_leader_changes_seen_total reports the leader changes. To rule out a slow disk and confirm that the disk is reasonably fast, 99th percentile of the etcd_disk_wal_fsync_duration_seconds_bucket should be less than 10ms. Query in metrics UI:

histogram_quantile(0.99, sum by (instance, le) (irate(etcd_disk_wal_fsync_duration_seconds_bucket{job="etcd"}[5m])))

Console dashboards #

In the OpenShift dashboard console under Observe section, select the etcd dashboard. There are both RPC rate as well as Disk Sync Duration dashboards which will assist with further issues.

Resource exhaustion #

It can happen that etcd responds slower due to CPU resource exhaustion. This was seen in some cases when one application was requesting too much CPU which led to this alert firing for multiple methods.

Often if this is the case, we also see etcd_disk_wal_fsync_duration_seconds_bucket slower as well.

To confirm this is the cause of the slow requests either:

  1. In OpenShift console on primary page under “Cluster utilization” view the requested CPU vs available.

  2. PromQL query is the following to see top consumers of CPU:

      topk(25, sort_desc(
        sum by (namespace) (
            sum(avg_over_time(pod:container_cpu_usage:sum{container="",pod!=""}[5m])) BY (namespace, pod)
            on(pod,namespace) group_left(node) (node_namespace_pod:kube_pod_info:)
          on(node) group_left(role) (max by (node) (kube_node_role{role=~".+"}))

Mitigation #

Fragmented state #

In the case of slow fisk or when the etcd DB size increases, we can defragment existing etcd DB to optimize DB consumption as described in etcdDefragmentation. Run the following command in all etcd pods.

$ etcdctl defrag

As validation, check the endpoint status of etcd members to know the reduced size of etcd DB. Use for this purpose the same diagnostic approaches as listed above. More space should be available now.

Further info on etcd best practices can be found in the etcdPractices.