etcd/Documentation/metrics.md

Metrics

NOTE: The metrics feature is considered as an experimental. We might add/change/remove metrics without warning in the future releases.

etcd uses Prometheus for metrics reporting in the server. The metrics can be used for real-time monitoring and debugging.

The simplest way to see the available metrics is to cURL the metrics endpoint /metrics of etcd. The format is described here.

You can also follow the doc here to start a Promethus server and monitor etcd metrics.

The naming of metrics follows the suggested best practice of Promethus. A metric name has an etcd prefix as its namespace and a subsystem prefix (for example wal and etcdserver).

etcd now exposes the following metrics:

etcdserver

Name	Description	Type
file_descriptors_used_total	The total number of file descriptors used	Gauge
proposal_durations_milliseconds	The latency distributions of committing proposal	Summary
pending_proposal_total	The total number of pending proposals	Gauge
proposal_failed_total	The total number of failed proposals	Counter

High file descriptors (file_descriptors_used_total) usage (near the file descriptors limitation of the process) indicates a potential out of file descriptors issue. That might cause etcd fails to create new WAL files and panics.

Proposal durations (proposal_durations_milliseconds) give you an summary about the proposal commit latency. Latency can be introduced into this process by network and disk IO.

Pending proposal (pending_proposal_total) gives you an idea about how many proposal are in the queue and waiting for commit. An increasing pending number indicates a high client load or an unstable cluster.

Failed proposals (proposal_failed_total) are normally related to two issues: temporary failures related to a leader election or longer duration downtime caused by a loss of quorum in the cluster.

store

These metrics describe the accesses into the data store of etcd members that exist in the cluster. They are useful to count what kind of actions are taken by users. It is also useful to see and whether all etcd members "see" the same set of data mutations, and whether reads and watches (which are local) are equally distributed.

All these metrics are prefixed with etcd_store_.

Name	Description	Type
reads_total	Total number of reads from store, should differ among etcd members (local reads).	Counter(action)
writes_total	Total number of writes to store, should be same among all etcd members.	Counter(action)
reads_failed_total	Number of failed reads from store (e.g. key missing) on local reads.	Counter(action)
writes_failed_total	Number of failed writes to store (e.g. failed compare and swap).	Counter(action)
expires_total	Total number of expired keys (due to TTL).	Counter
watch_requests_totals	Total number of incoming watch requests to this etcd member (local watches).	Counter
watchers	Current count of active watchers on this etcd member.	Gauge

Both reads_total and writes_total count both successful and failed requests. reads_failed_total and writes_failed_total count failed requests. A lot of failed writes indicate possible contentions on keys (e.g. when doing compareAndSet), and read failures indicate that some clients try to access keys that don't exist.

Example Prometheus queries that may be useful from these metrics (across all etcd members):

• sum(rate(etcd_store_reads_total{job="etcd"}[1m])) by (action) max(rate(etcd_store_writes_total{job="etcd"}[1m])) by (action)

  Rate of reads and writes by action, across all servers across a time window of 1m. The reason why max is used for writes as opposed to sum for reads is because all of etcd nodes in the cluster apply all writes to their stores. Shows the rate of successfull readonly/write queries across all servers, across a time window of 1m.

• sum(rate(etcd_store_watch_requests_total{job="etcd"}[1m]))

  Shows rate of new watch requests per second. Likely driven by how often watched keys change.

• sum(etcd_store_watchers{job="etcd"})

  Number of active watchers across all etcd servers.

wal

Name	Description	Type
fsync_durations_microseconds	The latency distributions of fsync called by wal	Summary
last_index_saved	The index of the last entry saved by wal	Gauge

Abnormally high fsync duration (fsync_durations_microseconds) indicates disk issues and might cause the cluster to be unstable.

snapshot

Name	Description	Type
snapshot_save_total_durations_microseconds	The total latency distributions of save called by snapshot	Summary

Abnormally high snapshot duration (snapshot_save_total_durations_microseconds) indicates disk issues and might cause the cluster to be unstable.

rafthttp

Name	Description	Type	Labels
message_sent_latency_microseconds	The latency distributions of messages sent	Summary	sendingType, msgType, remoteID
message_sent_failed_total	The total number of failed messages sent	Summary	sendingType, msgType, remoteID

Abnormally high message duration (message_sent_latency_microseconds) indicates network issues and might cause the cluster to be unstable.

An increase in message failures (message_sent_failed_total) indicates more severe network issues and might cause the cluster to be unstable.

Label sendingType is the connection type to send messages. message, msgapp and msgappv2 use HTTP streaming, while pipeline does HTTP request for each message.

Label msgType is the type of raft message. MsgApp is log replication message; MsgSnap is snapshot install message; MsgProp is proposal forward message; the others are used to maintain raft internal status. If you have a large snapshot, you would expect a long msgSnap sending latency. For other types of messages, you would expect low latency, which is comparable to your ping latency if you have enough network bandwidth.

Label remoteID is the member ID of the message destination.

proxy

etcd members operating in proxy mode do not do store operations. They forward all requests to cluster instances.

Tracking the rate of requests coming from a proxy allows one to pin down which machine is performing most reads/writes.

All these metrics are prefixed with etcd_proxy_

Name	Description	Type
requests_total	Total number of requests by this proxy instance. .	Counter(method)
handled_total	Total number of fully handled requests, with responses from etcd members.	Counter(method)
dropped_total	Total number of dropped requests due to forwarding errors to etcd members.	Counter(method,error)
handling_duration_seconds	Bucketed handling times by HTTP method, including round trip to member instances.	Histogram(method)

Example Prometheus queries that may be useful from these metrics (across all etcd servers):

• sum(rate(etcd_proxy_handled_total{job="etcd"}[1m])) by (method)

  Rate of requests (by HTTP method) handled by all proxies, across a window of 1m.

• histogram_quantile(0.9, sum(increase(etcd_proxy_events_handling_time_seconds_bucket{job="etcd",method="GET"}[5m])) by (le)) histogram_quantile(0.9, sum(increase(etcd_proxy_events_handling_time_seconds_bucket{job="etcd",method!="GET"}[5m])) by (le))

  Show the 0.90-tile latency (in seconds) of handling of user requestsacross all proxy machines, with a window of 5m.

• sum(rate(etcd_proxy_dropped_total{job="etcd"}[1m])) by (proxying_error)

  Number of failed request on the proxy. This should be 0, spikes here indicate connectivity issues to etcd cluster.