Optimised /api/v1/series for blocks storage

[pracucci]

What this PR does:
While working on #2794, I've compared our implementation to fetch series with Prometheus one and I've noticed Prometheus is easier. I've done the refactoring and written and benchmark and turned out it performs better too.

benchmark                                           old ns/op     new ns/op     delta
Benchmark_Ingester_v2MetricsForLabelMatchers-12     250057594     226011946     -9.62%

benchmark                                           old allocs     new allocs     delta
Benchmark_Ingester_v2MetricsForLabelMatchers-12     1104071        800196         -27.52%

benchmark                                           old bytes     new bytes     delta
Benchmark_Ingester_v2MetricsForLabelMatchers-12     62724828      51382654      -18.08%

Which issue(s) this PR fixes:
N/A

Checklist

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

[pstibrany]

Sets with series are not sorted, so merging them later doesn't work properly. This can be shown by using modified request in the benchmark/test:

	req := &client.MetricsForLabelMatchersRequest{
		StartTimestampMs: now,
		EndTimestampMs:   now,
		MatchersSet: []*client.LabelMatchers{
			{Matchers: []*client.LabelMatcher{
				{Type: client.REGEX_MATCH, Name: model.MetricNameLabel, Value: "test.*"},
			}},
			{Matchers: []*client.LabelMatcher{
				{Type: client.REGEX_MATCH, Name: model.MetricNameLabel, Value: "test.*0"}, // ending with 0
			}},
		},
	}

This should still return all numSeries series, but returns more of them.

Can you rerun benchmarks with sorting, and see whether it is still an improvement or not?

[pracucci]

Wow, thanks for spotting it! I copied the logic from Prometheus:
https://github.com/prometheus/prometheus/blob/b521612042ec87103088b45b7b5dfee3bb8dc732/web/api/v1/api.go#L604-L608

Either Prometheus /series API endpoint doesn't guarantee any deduplication, or there's a bug in Prometheus too. I will open an issue there.

Back to us, I modified the logic to use .Select(true, ...) (sorted) for now and re-run the benchmark (updated PR description). There's still some benefit, less then before.

[pracucci]

Opened this issue in Prometheus: prometheus/prometheus#7801. Will follow up this PR accordingly, if any change is required.

[pstibrany]

👍 Still looks like a nice improvement to me.