MQE: Add support for label_join

[jhesketh]

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.

[charleskorn]

Overall LGTM.

Could you please enable the existing label_join benchmark in pkg/streamingpromql/benchmarks/benchmarks.go and post the comparison between MQE and Prometheus' engine?

[charleskorn]

LGTM - could you please share the benchmark results?

[jhesketh]

Benchmark results

                                                                                    │  Prometheus  │               Mimir                │
                                                                                    │    sec/op    │    sec/op     vs base              │
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                    754.1µ ±  7%   708.3µ ±  1%  -6.08% (p=0.002 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8       784.2µ ± 14%   734.1µ ±  4%  -6.38% (p=0.002 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8      925.5µ ±  2%   866.2µ ±  1%  -6.41% (p=0.002 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                  2.459m ±  1%   2.436m ±  5%       ~ (p=0.589 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8     3.608m ±  1%   3.571m ±  1%  -1.01% (p=0.026 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8    14.29m ±  5%   14.17m ± 10%       ~ (p=0.589 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                 28.92m ±  2%   27.48m ±  2%  -4.98% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8    52.52m ±  1%   51.59m ±  3%       ~ (p=0.310 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8   251.5m ±  0%   245.0m ±  2%  -2.61% (p=0.004 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_instant_query-8                           760.4µ ±  4%   700.6µ ±  2%  -7.86% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8              742.5µ ±  5%   711.1µ ±  2%  -4.23% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8             888.1µ ±  1%   841.4µ ±  2%  -5.26% (p=0.002 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_instant_query-8                         2.397m ±  1%   2.353m ±  1%  -1.82% (p=0.009 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8            3.604m ±  1%   3.506m ± 11%       ~ (p=0.065 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8           14.19m ±  3%   14.37m ±  4%       ~ (p=0.240 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_instant_query-8                        27.75m ±  3%   26.92m ±  2%  -3.00% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8           52.13m ±  2%   50.50m ±  2%  -3.12% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8          249.4m ±  2%   246.3m ±  0%  -1.23% (p=0.002 n=6)
geomean                                                                               6.615m         6.396m        -3.31%

                                                                                    │  Prometheus  │                Mimir                │
                                                                                    │     B/op     │     B/op      vs base               │
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                    27.72Ki ± 0%   24.70Ki ± 0%  -10.91% (p=0.002 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8       28.82Ki ± 0%   25.43Ki ± 0%  -11.77% (p=0.002 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8      33.36Ki ± 0%   29.95Ki ± 0%  -10.20% (p=0.002 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                  184.3Ki ± 0%   163.4Ki ± 0%  -11.33% (p=0.002 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8     235.7Ki ± 0%   221.0Ki ± 0%   -6.20% (p=0.002 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8    659.4Ki ± 1%   644.1Ki ± 0%   -2.31% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                 3.165Mi ± 0%   2.825Mi ± 1%  -10.76% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8    4.165Mi ± 0%   4.016Mi ± 1%   -3.58% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8   12.41Mi ± 2%   12.05Mi ± 0%   -2.89% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_instant_query-8                           24.93Ki ± 0%   21.88Ki ± 0%  -12.23% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8              25.99Ki ± 0%   22.63Ki ± 0%  -12.92% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8             30.52Ki ± 0%   27.07Ki ± 0%  -11.33% (p=0.002 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_instant_query-8                         175.9Ki ± 0%   156.8Ki ± 0%  -10.87% (p=0.002 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8            227.1Ki ± 0%   214.3Ki ± 0%   -5.63% (p=0.002 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8           649.9Ki ± 1%   637.8Ki ± 1%   -1.87% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_instant_query-8                        3.073Mi ± 0%   2.792Mi ± 0%   -9.17% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8           4.074Mi ± 0%   3.983Mi ± 0%   -2.24% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8          12.12Mi ± 2%   12.02Mi ± 0%   -0.86% (p=0.002 n=6)
geomean                                                                               361.7Ki        333.8Ki        -7.71%

                                                                                    │ Prometheus  │               Mimir                │
                                                                                    │  allocs/op  │  allocs/op   vs base               │
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                     514.0 ± 0%    422.0 ± 0%  -17.90% (p=0.002 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8        524.0 ± 0%    429.0 ± 0%  -18.13% (p=0.002 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8       550.0 ± 0%    455.0 ± 0%  -17.27% (p=0.002 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                  2.824k ± 0%   2.724k ± 0%   -3.54% (p=0.002 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8     3.236k ± 0%   3.135k ± 0%   -3.14% (p=0.002 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8    5.852k ± 0%   5.750k ± 0%   -1.73% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                 47.03k ± 0%   46.90k ± 0%   -0.27% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8    55.07k ± 0%   54.97k ± 0%   -0.19% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8   107.2k ± 0%   107.0k ± 0%   -0.12% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_instant_query-8                            473.0 ± 0%    381.0 ± 0%  -19.45% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8               483.0 ± 0%    388.0 ± 0%  -19.67% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8              509.0 ± 0%    414.0 ± 0%  -18.66% (p=0.002 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_instant_query-8                         2.585k ± 0%   2.493k ± 0%   -3.56% (p=0.002 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8            2.997k ± 0%   2.903k ± 0%   -3.14% (p=0.002 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8           5.613k ± 0%   5.519k ± 0%   -1.67% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_instant_query-8                        42.99k ± 0%   43.87k ± 0%   +2.04% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8           51.02k ± 0%   51.93k ± 0%   +1.77% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8          103.1k ± 0%   104.0k ± 0%   +0.89% (p=0.002 n=6)
geomean                                                                               4.886k        4.530k        -7.27%

                                                                                    │  Prometheus  │               Mimir                │
                                                                                    │      B       │      B        vs base              │
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                    68.58Mi ± 2%   68.63Mi ± 1%       ~ (p=0.937 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8       68.52Mi ± 2%   67.93Mi ± 1%       ~ (p=0.589 n=6)
Query/label_replace(a_1,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8      68.10Mi ± 1%   67.60Mi ± 1%  -0.73% (p=0.022 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                  65.83Mi ± 1%   65.45Mi ± 2%       ~ (p=0.240 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8     66.33Mi ± 1%   66.26Mi ± 1%       ~ (p=0.699 n=6)
Query/label_replace(a_100,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8    69.04Mi ± 1%   69.14Mi ± 1%       ~ (p=0.485 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_instant_query-8                 67.40Mi ± 1%   67.23Mi ± 1%       ~ (p=0.937 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_100_steps-8    73.86Mi ± 1%   76.00Mi ± 1%  +2.91% (p=0.002 n=6)
Query/label_replace(a_2000,_'l2',_'$1',_'l',_'(.*)'),_range_query_with_1000_steps-8   126.0Mi ± 3%   130.2Mi ± 1%  +3.34% (p=0.002 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_instant_query-8                           68.84Mi ± 1%   68.88Mi ± 2%       ~ (p=0.699 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8              69.64Mi ± 2%   68.68Mi ± 1%       ~ (p=0.065 n=6)
Query/label_join(a_1,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8             68.47Mi ± 2%   68.22Mi ± 2%       ~ (p=0.589 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_instant_query-8                         65.70Mi ± 1%   65.72Mi ± 1%       ~ (p=0.818 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8            65.99Mi ± 1%   66.22Mi ± 1%       ~ (p=0.589 n=6)
Query/label_join(a_100,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8           68.96Mi ± 1%   68.94Mi ± 1%       ~ (p=0.589 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_instant_query-8                        67.35Mi ± 3%   67.76Mi ± 1%       ~ (p=0.132 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_100_steps-8           74.35Mi ± 1%   76.10Mi ± 2%  +2.35% (p=0.002 n=6)
Query/label_join(a_2000,_'l2',_'-',_'l',_'l'),_range_query_with_1000_steps-8          129.3Mi ± 2%   129.6Mi ± 1%       ~ (p=0.394 n=6)
geomean                                                                               73.42Mi        73.65Mi       +0.31%

tldr: >10% better memory, 3-5% better CPU

[charleskorn]

I'm surprised to see that MQE allocates more often than Prometheus' engine for label_join over many series - any idea why that is? (I'm wondering if we've missed something.)