This reverts commit 5a3abfa041.
Reason for revert: exemplars aren't in wide use because they have numerous issues which prevent their adoption (see below).
Adding support for examplars into VictoriaMetrics introduces non-trivial code changes. These code changes need to be supported forever
once the release of VictoriaMetrics with exemplar support is published. That's why I don't think this is a good feature despite
that the source code of the reverted commit has an excellent quality. See https://docs.victoriametrics.com/goals/ .
Issues with Prometheus exemplars:
- Prometheus still has only experimental support for exemplars after more than three years since they were introduced.
It stores exemplars in memory, so they are lost after Prometheus restart. This doesn't look like production-ready feature.
See 0a2f3b3794/content/docs/instrumenting/exposition_formats.md (L153-L159)
and https://prometheus.io/docs/prometheus/latest/feature_flags/#exemplars-storage
- It is very non-trivial to expose exemplars alongside metrics in your application, since the official Prometheus SDKs
for metrics' exposition ( https://prometheus.io/docs/instrumenting/clientlibs/ ) either have very hard-to-use API
for exposing histograms or do not have this API at all. For example, try figuring out how to expose exemplars
via https://pkg.go.dev/github.com/prometheus/client_golang@v1.19.1/prometheus .
- It looks like exemplars are supported for Histogram metric types only -
see https://pkg.go.dev/github.com/prometheus/client_golang@v1.19.1/prometheus#Timer.ObserveDurationWithExemplar .
Exemplars aren't supported for Counter, Gauge and Summary metric types.
- Grafana has very poor support for Prometheus exemplars. It looks like it supports exemplars only when the query
contains histogram_quantile() function. It queries exemplars via special Prometheus API -
https://prometheus.io/docs/prometheus/latest/querying/api/#querying-exemplars - (which is still marked as experimental, btw.)
and then displays all the returned exemplars on the graph as special dots. The issue is that this doesn't work
in production in most cases when the histogram_quantile() is calculated over thousands of histogram buckets
exposed by big number of application instances. Every histogram bucket may expose an exemplar on every timestamp shown on the graph.
This makes the graph unusable, since it is litterally filled with thousands of exemplar dots.
Neither Prometheus API nor Grafana doesn't provide the ability to filter out unneeded exemplars.
- Exemplars are usually connected to traces. While traces are good for some
I doubt exemplars will become production-ready in the near future because of the issues outlined above.
Alternative to exemplars:
Exemplars are marketed as a silver bullet for the correlation between metrics, traces and logs -
just click the exemplar dot on some graph in Grafana and instantly see the corresponding trace or log entry!
This doesn't work as expected in production as shown above. Are there better solutions, which work in production?
Yes - just use time-based and label-based correlation between metrics, traces and logs. Assign the same `job`
and `instance` labels to metrics, logs and traces, so you can quickly find the needed trace or log entry
by these labes on the time range with the anomaly on metrics' graph.
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/pull/5982
This code adds Exemplars to VMagent and the promscrape parser adhering
to OpenMetrics Specifications. This will allow forwarding of exemplars
to Prometheus and other third party apps that support OpenMetrics specs.
---------
Signed-off-by: Ted Possible <ted_possible@cable.comcast.com>
(cherry picked from commit 5a3abfa041)
- Add Try* prefix to functions, which return bool result in order to improve readability and reduce the probability of missing check
for the result returned from these functions.
- Call the adjustSampleValues() only once on input samples. Previously it was called on every attempt to flush data to peristent queue.
- Properly restore the initial state of WriteRequest passed to tryPushWriteRequest() before returning from this function
after unsuccessful push to persistent queue. Previously a part of WriteRequest samples may be lost in such case.
- Add -remoteWrite.dropSamplesOnOverload command-line flag, which can be used for dropping incoming samples instead
of returning 429 Too Many Requests error to the client when -remoteWrite.disableOnDiskQueue is set and the remote storage
cannot keep up with the data ingestion rate.
- Add vmagent_remotewrite_samples_dropped_total metric, which counts the number of dropped samples.
- Add vmagent_remotewrite_push_failures_total metric, which counts the number of unsuccessful attempts to push
data to persistent queue when -remoteWrite.disableOnDiskQueue is set.
- Remove vmagent_remotewrite_aggregation_metrics_dropped_total and vm_promscrape_push_samples_dropped_total metrics,
because they are replaced with vmagent_remotewrite_samples_dropped_total metric.
- Update 'Disabling on-disk persistence' docs at docs/vmagent.md
- Update stale comments in the code
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/pull/5088
Updates https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2110
* app/vmagent: allow to disabled on-disk queue
Previously, it wasn't possible to build data processing pipeline with a
chain of vmagents. In case when remoteWrite for the last vmagent in the
chain wasn't accessible, it persisted data only when it has enough disk
capacity. If disk queue is full, it started to silently drop ingested
metrics.
New flags allows to disable on-disk persistent and immediatly return an
error if remoteWrite is not accessible anymore. It blocks any writes and
notify client, that data ingestion isn't possible.
Main use case for this feature - use external queue such as kafka for
data persistence.
https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2110
* adds test, updates readme
* apply review suggestions
* update docs for vmagent
* makes linter happy
---------
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
The standard Snappy encoder from github.com/golang/snappy shows quite good performance number
for compressing the Prometheus remote_write proto messages according to the added benchmarks,
so there is no need in switching to github.com/klauspost/compress/s2 yet.
