rafthttp: add Raft send latency metric for writes #10023
Conversation
|
Address #9438. |
Codecov Report
@@ Coverage Diff @@
## master #10023 +/- ##
==========================================
- Coverage 71.59% 71.51% -0.09%
==========================================
Files 390 390
Lines 36258 36265 +7
==========================================
- Hits 25960 25935 -25
- Misses 8488 8519 +31
- Partials 1810 1811 +1
Continue to review full report at Codecov.
|
etcdserver/api/rafthttp/stream.go
Outdated
| // heartbeats are tracked via prober https://github.com/coreos/etcd/pull/10022 | ||
| if m.Type == raftpb.MsgProp || m.Type == raftpb.MsgApp || m.Type == raftpb.MsgAppResp { | ||
| raftSendSeconds.WithLabelValues(m.Type.String(), types.ID(m.To).String()).Observe(took.Seconds()) | ||
| } |
There was a problem hiding this comment.
Is there a reason that MsgProp, MsgApp and MsgAppResp are chosen here? I was reading the doc, and you explained snapshot and heartbeats, how about other message types?
| // highest bucket start of 0.0001 sec * 2^15 == 3.2768 sec | ||
| Buckets: prometheus.ExponentialBuckets(0.0001, 2, 16), | ||
| }, | ||
| []string{"Type", "To"}, |
There was a problem hiding this comment.
Do you mean []string{"To"}?
Signed-off-by: Gyuho Lee <leegyuho@amazon.com>
…econds" metric
Currently, only v2 metrics ("stats.FollowerStats") tracks Raft message
send latencies. Add Prometheus histogram to track Raft messages for
writes, since heartbeats are probed (see etcd-io#10022)
and snapshots are already being tracked via etcd-io#9997.
```
etcd_network_raft_send_total_duration_seconds_bucket{To="7339c4e5e833c029",Type="MsgProp",le="0.0001"} 1
etcd_network_raft_send_total_duration_seconds_bucket{To="7339c4e5e833c029",Type="MsgProp",le="0.0002"} 1
etcd_network_raft_send_total_duration_seconds_bucket{To="729934363faa4a24",Type="MsgApp",le="0.0001"} 9
etcd_network_raft_send_total_duration_seconds_bucket{To="729934363faa4a24",Type="MsgApp",le="0.0002"} 9
etcd_network_raft_send_total_duration_seconds_bucket{To="7339c4e5e833c029",Type="MsgAppResp",le="0.0001"} 8
etcd_network_raft_send_total_duration_seconds_bucket{To="7339c4e5e833c029",Type="MsgAppResp",le="0.0002"} 8
```
Signed-off-by: Gyuho Lee <leegyuho@amazon.com>
Those are most frequent ones that would give us enough sample size to investigate any network issues. Other important ones are heartbeat and v3 snapshot messages:
By adding "From" field, we can just use one metrics to track multiple message types. Can you confirm with GKE team, if this does not break anything in production? I will check with other production users. |
|
Oh, this is a new metrics, so should not break anything when backported. |
|
@gyuho I do need to check if it's safe to add label though. Checking now, get back to you soon. |
| // snapshot sends are tracked via separate metrics https://github.com/etcd-io/etcd/pull/9997 | ||
| // heartbeats are tracked via prober https://github.com/etcd-io/etcd/pull/10022 | ||
| // TODO: track other messages? | ||
| if m.Type == raftpb.MsgProp || |
There was a problem hiding this comment.
this is not the sending time (flushing the buffer to the network card). we actually send the messages at line 211 when flush is called explicitly.
There was a problem hiding this comment.
i would rather just remove this metrics. it is not easy to measure after all with buffers all over the places.
There was a problem hiding this comment.
@xiang90 Good point. For now, tracking snapshots and heartbeats is good enough. Will come back to this later.
|
Will get back to this later. |
Currently, only v2 metrics ("stats.FollowerStats") tracks Raft message
send latencies. Add Prometheus histogram to track Raft messages for
writes, since heartbeats are probed (see #10022)
and snapshots are already being tracked via #9997.