Add query per stage latency metrics - #839
Conversation
…er_stage_latency_metrics
| } | ||
| searchContext.getResponseBuilder().setDiagnostics(diagnostics); | ||
|
|
||
| // TODO: These are the diagnostics I want to publish to prometheus, I'll try to figure out |
There was a problem hiding this comment.
I just left these comments here as a reference for myself.
| .observe(entry.getValue()); | ||
| } | ||
| searchStageLatencyMs.labelValues(index, "rescore").observe(diagnostics.getRescoreTimeMs());//adding extra rescore metric to avoid calculating average of all rescorers | ||
| for (Map.Entry<String, Double> entry : diagnostics.getRescorersTimeMsMap().entrySet()) { |
There was a problem hiding this comment.
For now, searchStageLatencyMs only has the rescorer latency for each rescorer, but for an initial overview, a general latency for all rescorers might also be useful.
| @@ -151,7 +155,6 @@ public MetricSnapshots collect() { | |||
| if (publishVerboseMetrics) { | |||
There was a problem hiding this comment.
When would we set this to true?
There was a problem hiding this comment.
Only when we need to the extra metrics to investigate something
| metrics.add(searchTerminatedEarlyCount.collect()); | ||
| metrics.add(searchStageLatencyMs.collect());// Just adding this here should mean it gets published to prometheus, is that what we want? | ||
| // when is publishVerboseMetrics set to true, I couldn't find this metric in the grafana shard without any filters? | ||
| // maybe it makes sense to add an extra parameter like publishSearchStageLatencyMs with default value true (?) |
There was a problem hiding this comment.
Does it make sense to have an option to turn this on and off separately?
There was a problem hiding this comment.
There is a live index setting which enables publishing verbose metrics
There was a problem hiding this comment.
I think if a metric is per-hit, it should come under the verbose metrics
There was a problem hiding this comment.
Yes, that makes sense. I was considering adding something like publishOnlyPerQueryStageLatency, which could offer a way to only publish searchStageLatencyMs without searchResponseSizeBytes and searchResponseTotalHits. (All three would still be published if verbose is set to true.)
But if we only turn this on in case we want to investigate something, or the performance impact is not too high, I suppose that wouldn't be necessary.
| @@ -328,7 +328,8 @@ public SearchResponse handle(IndexState indexState, SearchRequest searchRequest) | |||
| InnerHitFetchTask::getDiagnostic))); | |||
| } | |||
| searchContext.getResponseBuilder().setDiagnostics(diagnostics); | |||
There was a problem hiding this comment.
I think the diagnostics object here contains all of the required metrics, including the rescorer metrics. Can you please confirm once?
There was a problem hiding this comment.
Yes, it looks like it here. What I couldn't find there was a total rescore time metric, but just adding up the time of all rescorers in a prometheus query might be an easier option.
2 participants
Working on publishing per stage latency metrics via prometheus.