Check cached postings TTL before returning from cache + metrics

[pracucci]

In this PR I propose two changes to PostingsForMatchers cache:

1. Check if the TTL for cached postings is still valid before returning it from cache, to fix a race condition that could happen between a goroutine running expire() and another one skipping the expire() execution because it's already in progress.
2. Add metrics to PostingsForMatchers cache, to have better visibility over it. This is something I wanted to do since a long time. The design I picked is to allow to pass the metrics struct as DB options, so that we can use 1 single struct for all per-tenant TSDBs in a Mimir ingester.

The following benchmark shows the difference betweeen:

• 01bb37aae: the commit before Reduce PostingsForMatchersCache.expire() pressure on mutex #734
• 6c2603082: the commit at Reduce PostingsForMatchersCache.expire() pressure on mutex #734
• main
• This PR

goos: darwin
goarch: arm64
pkg: github.com/prometheus/prometheus/tsdb
cpu: Apple M3 Pro
                                                          │ 01bb37aae.txt │           6c2603082.txt            │              main.txt              │               pr.txt               │
                                                          │    sec/op     │   sec/op     vs base               │   sec/op     vs base               │   sec/op     vs base               │
PostingsForMatchersCache/no_evictions-11                      594.8n ± 3%   584.2n ± 1%   -1.78% (p=0.015 n=6)   602.9n ± 3%        ~ (p=0.589 n=6)   644.6n ± 1%   +8.38% (p=0.002 n=6)
PostingsForMatchersCache/high_eviction_rate-11                10.52µ ± 5%   10.39µ ± 1%   -1.28% (p=0.002 n=6)   10.85µ ± 1%        ~ (p=0.065 n=6)   10.77µ ± 0%        ~ (p=0.065 n=6)
PostingsForMatchersCache_ConcurrencyOnHighEvictionRate-11    1411.5n ± 2%   301.1n ± 1%  -78.67% (p=0.002 n=6)   306.2n ± 1%  -78.31% (p=0.002 n=6)   331.2n ± 2%  -76.54% (p=0.002 n=6)
geomean                                                       2.067µ        1.222µ       -40.86%                 1.260µ       -39.03%                 1.320µ       -36.15%

                                                          │ 01bb37aae.txt │             6c2603082.txt             │               main.txt                │               pr.txt                │
                                                          │     B/op      │     B/op      vs base                 │     B/op      vs base                 │     B/op      vs base               │
PostingsForMatchersCache/no_evictions-11                       958.0 ± 0%     958.0 ± 0%        ~ (p=1.000 n=6) ¹     958.0 ± 0%        ~ (p=1.000 n=6) ¹     974.0 ± 0%   +1.67% (p=0.002 n=6)
PostingsForMatchersCache/high_eviction_rate-11               26.38Ki ± 0%   26.38Ki ± 0%        ~ (p=1.000 n=6) ¹   26.38Ki ± 0%        ~ (p=1.000 n=6) ¹   26.32Ki ± 0%   -0.24% (p=0.002 n=6)
PostingsForMatchersCache_ConcurrencyOnHighEvictionRate-11     1441.0 ± 0%    1017.0 ± 0%  -29.42% (p=0.002 n=6)      1014.0 ± 0%  -29.63% (p=0.002 n=6)      1024.5 ± 0%  -28.90% (p=0.002 n=6)
geomean                                                      3.263Ki        2.905Ki       -10.97%                   2.902Ki       -11.05%                   2.926Ki       -10.33%
¹ all samples are equal

                                                          │ 01bb37aae.txt │            6c2603082.txt            │              main.txt               │               pr.txt                │
                                                          │   allocs/op   │ allocs/op   vs base                 │ allocs/op   vs base                 │ allocs/op   vs base                 │
PostingsForMatchersCache/no_evictions-11                       20.00 ± 0%   20.00 ± 0%        ~ (p=1.000 n=6) ¹   20.00 ± 0%        ~ (p=1.000 n=6) ¹   20.00 ± 0%        ~ (p=1.000 n=6) ¹
PostingsForMatchersCache/high_eviction_rate-11                 48.00 ± 0%   48.00 ± 0%        ~ (p=1.000 n=6) ¹   48.00 ± 0%        ~ (p=1.000 n=6) ¹   46.00 ± 0%   -4.17% (p=0.002 n=6)
PostingsForMatchersCache_ConcurrencyOnHighEvictionRate-11      29.00 ± 0%   21.00 ± 0%  -27.59% (p=0.002 n=6)     21.00 ± 0%  -27.59% (p=0.002 n=6)     20.00 ± 5%  -31.03% (p=0.002 n=6)
geomean                                                        30.31        27.22       -10.20%                   27.22       -10.20%                   26.40       -12.89%
¹ all samples are equal

[charleskorn]

Approach LGTM.

It'd be interesting to benchmark this and compare the cost of checking the done channel with the approach prior to #734.

[charleskorn]

(not something for this PR) Metrics for the PFMC in general would be handy - would be nice to see the rate of cache hits and misses, as well as the rate of cache evictions and their reason (TTL vs cache growing too large).

[pracucci]

as well as the rate of cache evictions and their reason (TTL vs cache growing too large)

Will do in a separate PR.

[charleskorn]

Overall LGTM

[charleskorn]

Would be good to include evaluationCompletedAt here as well, to replace what was reverted from #820.

[pracucci]

Unfortunately we can't do it, because there's no guarantee the promise already completed at this step and so it's not safe to access evaluationCompletedAt

[pracucci]

I added the timestamp in another span tho: 6da4d10

[charleskorn]

I think we need to find a way to log evaluationCompletedAt when we use a cached entry, but doing this doesn't need to block this PR.

[pracucci]

find a way to log evaluationCompletedAt when we use a cached entry

Added to the issue: https://github.com/grafana/pir-actions/issues/307