POC: Cost Attribution Proposal 1.2

[ying-jeanne]

What this PR does

Which issue(s) this PR fixes or relates to

Fixes #

Checklist

• Tests updated.
• Documentation added.
• CHANGELOG.md updated - the order of entries should be [CHANGE], [FEATURE], [ENHANCEMENT], [BUGFIX].
• about-versioning.md updated with experimental features.

[ying-jeanne]

I think here we might still want ts, when we decrement a active series counter that are "active" we may want to keep the timestamp updated.

[colega]

Why would we want to update the timestamp of something that didn't have activity?

[colega]

I reviewed mostly the existing code changes, skipping the costattribution package implementation for now.

Please ping me back when this is addressed.

Meanwhile, while I didn't look into Tracker implementation, since it's a hot path, I wonder if we could replace the prometheus.Gauge implementation by a custom map. Of course we'll need to benchmark this.

[colega]

I think we don't need an interface here, it just makes things more complex and slow. Now we can't compare two trackers by just comparing the pointers, and we expose too many internal details on exported method just to make that comparison.

There are only two implementations of Tracker: the one that does things and the one that is disabled, we can make the disabled one be the nil pointer of TrackerImpl (please rename that back to Tracker).

[ying-jeanne]

removed the interface, and it checks for nil pointer in the receiver instead. commit 7f0b372

[colega]

Since we have ConfigDiffers now, we don't need CurrentConfig or CurrentCostAttributionTracker methods, and we don't need to take mutex twice for them.

[ying-jeanne]

addressed in this commit 698a5c6#diff-74600e3203a9faaa94f7231d161362da3e3223aa024eb324b903e5fc27bb9f3dR116

[ying-jeanne]

Based on the benchmark results, the successfully pushed case shows a significant increase in allocations, which is expected due to tracking counters per cost label value, previously there is only one counter for received sample tracking. In discarded cases, differences are minimal since counters per reason were already tracked, with a notable percentage of memory consumption. The additional allocations primarily stem from tracking cost attribution-specific data.

BenchmarkDistributor_Push/all_samples_successfully_pushed-8      10           2432680 ns/op          212628 B/op         98 allocs/op
BenchmarkDistributor_Push/all_samples_successfully_pushed_with_cost_attribution_enabled-8    10           4021089 ns/op          287027 B/op        3110 allocs/op
BenchmarkDistributor_Push/too_many_labels_limit_reached-8          10          12637491 ns/op         1218316 B/op        5069 allocs/op
BenchmarkDistributor_Push/too_many_labels_limit_reached_with_cost_attribution_enabled-8         10           8159270 ns/op         1351149 B/op        9080 allocs/op
BenchmarkDistributor_Push/all_samples_go_to_metric_relabel_configs-8         10           4029890 ns/op         1414630 B/op       7099 allocs/op
BenchmarkDistributor_Push/max_label_value_length_limit_reached-8           10           7136839 ns/op         1234532 B/op        6069 allocs/op
BenchmarkDistributor_Push/timestamp_too_new-8          10           2085542 ns/op          345832 B/op        4069 allocs/op
BenchmarkDistributor_Push/ingestion_rate_limit_reached-8           10           1061883 ns/op           21572 B/op          59 allocs/op
BenchmarkDistributor_Push/max_label_name_length_limit_reached-8        10           4094097 ns/op         1150353 B/op        5067 allocs/op

[ying-jeanne]

The result of ingester push benchmark, the impact is less compare to distributor.

BenchmarkIngesterPush/ingester_push_succeeded-8                       10             49926 ns/op            9944 B/op         37 allocs/op
BenchmarkIngesterPush/ingester_push_succeeded_with_cost_attribution_enabled-8                 10             52132 ns/op           10514 B/op         51 allocs/op