Mimir query engine: avoid copying native histograms where possible

[charleskorn]

What this PR does

This PR improves the performance of queries over native histograms in MQE. It does this by avoiding creating copies of FloatHistograms where possible.

This improves query latency for queries over native histograms in MQE by up to 20%. This has no noticeable impact on queries over only floats, and makes MQE up to 50% faster than Prometheus' engine in some cases.

For example, previously, while evaluating a sum over native histograms, we would always copy FloatHistogram instances that were the first value for an output group. However, this is not necessary: we only need to make a copy if the same FloatHistogram will be used as the first value for subsequent points.

Performance comparison of this PR vs Prometheus' engine

goos: darwin
goarch: arm64
pkg: github.com/grafana/mimir/pkg/streamingpromql/benchmarks
                                                                 │      Prometheus      │                Mimir                │
                                                                 │        sec/op        │    sec/op     vs base               │
Query/nh_1,_instant_query                                                  158.8µ ±  8%   140.2µ ±  2%  -11.68% (p=0.002 n=6)
Query/nh_1,_range_query_with_100_steps                                     221.2µ ±  6%   192.1µ ±  2%  -13.15% (p=0.002 n=6)
Query/nh_1,_range_query_with_1000_steps                                    774.4µ ±  2%   637.9µ ±  1%  -17.63% (p=0.002 n=6)
Query/nh_100,_instant_query                                                2.118m ±  8%   1.946m ±  3%   -8.10% (p=0.009 n=6)
Query/nh_100,_range_query_with_100_steps                                   7.724m ± 18%   6.305m ±  3%  -18.38% (p=0.041 n=6)
Query/nh_100,_range_query_with_1000_steps                                  55.77m ±  1%   44.35m ±  6%  -20.48% (p=0.002 n=6)
Query/nh_2000,_instant_query                                               38.62m ± 14%   32.80m ±  7%  -15.08% (p=0.004 n=6)
Query/nh_2000,_range_query_with_100_steps                                  139.8m ±  1%   114.0m ± 12%  -18.45% (p=0.002 n=6)
Query/nh_2000,_range_query_with_1000_steps                                1031.3m ±  1%   809.3m ±  1%  -21.52% (p=0.002 n=6)
Query/sum(nh_1),_instant_query                                             145.0µ ±  5%   140.8µ ±  6%   -2.91% (p=0.041 n=6)
Query/sum(nh_1),_range_query_with_100_steps                                225.9µ ±  0%   195.3µ ±  1%  -13.52% (p=0.002 n=6)
Query/sum(nh_1),_range_query_with_1000_steps                               862.6µ ±  3%   650.6µ ±  3%  -24.57% (p=0.002 n=6)
Query/sum(nh_100),_instant_query                                           1.986m ±  2%   1.943m ±  1%   -2.16% (p=0.002 n=6)
Query/sum(nh_100),_range_query_with_100_steps                              7.069m ±  1%   6.813m ±  0%   -3.62% (p=0.002 n=6)
Query/sum(nh_100),_range_query_with_1000_steps                             55.44m ±  1%   48.88m ±  1%  -11.83% (p=0.002 n=6)
Query/sum(nh_2000),_instant_query                                          33.83m ±  1%   32.70m ±  1%   -3.36% (p=0.002 n=6)
Query/sum(nh_2000),_range_query_with_100_steps                             170.3m ±  1%   126.3m ±  1%  -25.87% (p=0.002 n=6)
Query/sum(nh_2000),_range_query_with_1000_steps                           1546.6m ±  1%   929.0m ±  1%  -39.93% (p=0.002 n=6)
Query/sum_by_(l)(nh_1),_instant_query                                      146.1µ ±  4%   140.1µ ±  1%   -4.08% (p=0.002 n=6)
Query/sum_by_(l)(nh_1),_range_query_with_100_steps                         232.3µ ±  3%   193.8µ ±  3%  -16.56% (p=0.002 n=6)
Query/sum_by_(l)(nh_1),_range_query_with_1000_steps                        859.1µ ±  1%   636.1µ ±  1%  -25.96% (p=0.002 n=6)
Query/sum_by_(l)(nh_100),_instant_query                                    2.032m ±  1%   1.959m ±  0%   -3.59% (p=0.002 n=6)
Query/sum_by_(l)(nh_100),_range_query_with_100_steps                       8.550m ±  1%   6.377m ±  0%  -25.42% (p=0.002 n=6)
Query/sum_by_(l)(nh_100),_range_query_with_1000_steps                      68.18m ±  2%   44.88m ±  1%  -34.16% (p=0.002 n=6)
Query/sum_by_(l)(nh_2000),_instant_query                                   35.04m ±  1%   34.18m ±  1%   -2.45% (p=0.002 n=6)
Query/sum_by_(l)(nh_2000),_range_query_with_100_steps                      198.4m ±  1%   115.3m ±  2%  -41.89% (p=0.002 n=6)
Query/sum_by_(l)(nh_2000),_range_query_with_1000_steps                    1759.5m ±  1%   819.6m ±  4%  -53.42% (p=0.002 n=6)

                                                                 │      Prometheus      │                Mimir                │
                                                                 │          B           │      B        vs base               │
Query/nh_1,_instant_query                                                  78.95Mi ± 2%   79.27Mi ± 1%        ~ (p=0.699 n=6)
Query/nh_1,_range_query_with_100_steps                                     72.15Mi ± 1%   71.22Mi ± 1%   -1.29% (p=0.002 n=6)
Query/nh_1,_range_query_with_1000_steps                                    70.04Mi ± 1%   70.20Mi ± 2%        ~ (p=0.699 n=6)
Query/nh_100,_instant_query                                                66.00Mi ± 1%   65.48Mi ± 0%   -0.78% (p=0.002 n=6)
Query/nh_100,_range_query_with_100_steps                                   70.69Mi ± 1%   70.78Mi ± 1%        ~ (p=0.416 n=6)
Query/nh_100,_range_query_with_1000_steps                                  117.8Mi ± 1%   118.0Mi ± 1%        ~ (p=0.485 n=6)
Query/nh_2000,_instant_query                                               68.50Mi ± 1%   67.95Mi ± 2%        ~ (p=0.093 n=6)
Query/nh_2000,_range_query_with_100_steps                                  173.9Mi ± 1%   175.8Mi ± 1%   +1.09% (p=0.015 n=6)
Query/nh_2000,_range_query_with_1000_steps                                 1.063Gi ± 2%   1.090Gi ± 3%        ~ (p=0.065 n=6)
Query/sum(nh_1),_instant_query                                             79.20Mi ± 1%   78.90Mi ± 1%        ~ (p=0.310 n=6)
Query/sum(nh_1),_range_query_with_100_steps                                71.27Mi ± 1%   72.12Mi ± 1%   +1.19% (p=0.026 n=6)
Query/sum(nh_1),_range_query_with_1000_steps                               70.38Mi ± 1%   70.33Mi ± 1%        ~ (p=0.615 n=6)
Query/sum(nh_100),_instant_query                                           66.10Mi ± 1%   65.51Mi ± 1%        ~ (p=0.056 n=6)
Query/sum(nh_100),_range_query_with_100_steps                              71.01Mi ± 1%   65.30Mi ± 1%   -8.04% (p=0.002 n=6)
Query/sum(nh_100),_range_query_with_1000_steps                            118.10Mi ± 1%   67.77Mi ± 1%  -42.61% (p=0.002 n=6)
Query/sum(nh_2000),_instant_query                                          68.60Mi ± 1%   66.36Mi ± 1%   -3.27% (p=0.002 n=6)
Query/sum(nh_2000),_range_query_with_100_steps                            172.78Mi ± 2%   66.95Mi ± 1%  -61.25% (p=0.002 n=6)
Query/sum(nh_2000),_range_query_with_1000_steps                           718.55Mi ± 4%   74.10Mi ± 1%  -89.69% (p=0.002 n=6)
Query/sum_by_(l)(nh_1),_instant_query                                      78.72Mi ± 2%   78.98Mi ± 1%        ~ (p=0.909 n=6)
Query/sum_by_(l)(nh_1),_range_query_with_100_steps                         71.11Mi ± 1%   71.42Mi ± 1%        ~ (p=0.260 n=6)
Query/sum_by_(l)(nh_1),_range_query_with_1000_steps                        70.97Mi ± 1%   70.41Mi ± 1%        ~ (p=0.093 n=6)
Query/sum_by_(l)(nh_100),_instant_query                                    65.93Mi ± 1%   65.91Mi ± 1%        ~ (p=0.818 n=6)
Query/sum_by_(l)(nh_100),_range_query_with_100_steps                       76.09Mi ± 0%   71.05Mi ± 1%   -6.62% (p=0.002 n=6)
Query/sum_by_(l)(nh_100),_range_query_with_1000_steps                      169.9Mi ± 1%   118.2Mi ± 1%  -30.40% (p=0.002 n=6)
Query/sum_by_(l)(nh_2000),_instant_query                                   70.18Mi ± 1%   69.38Mi ± 1%   -1.15% (p=0.024 n=6)
Query/sum_by_(l)(nh_2000),_range_query_with_100_steps                      280.8Mi ± 2%   177.7Mi ± 1%  -36.74% (p=0.002 n=6)
Query/sum_by_(l)(nh_2000),_range_query_with_1000_steps                     1.110Gi ± 0%   1.093Gi ± 5%        ~ (p=1.000 n=6)

Performance comparison of main vs this PR

Peak memory consumption is unchanged within the margin of error for the benchmarks.

goos: darwin
goarch: arm64
pkg: github.com/grafana/mimir/pkg/streamingpromql/benchmarks
                                                                          │   main.txt    │       with-copy-avoidance.txt       │
                                                                          │    sec/op     │    sec/op     vs base               │
Query/nh_1,_instant_query/Mimir-10                                           140.8µ ±  2%   140.2µ ±  2%        ~ (p=0.258 n=6)
Query/nh_1,_range_query_with_100_steps/Mimir-10                              193.1µ ±  2%   192.1µ ±  2%        ~ (p=0.394 n=6)
Query/nh_1,_range_query_with_1000_steps/Mimir-10                             639.6µ ±  2%   637.9µ ±  1%        ~ (p=0.937 n=6)
Query/nh_100,_instant_query/Mimir-10                                         1.949m ±  1%   1.946m ±  3%        ~ (p=0.818 n=6)
Query/nh_100,_range_query_with_100_steps/Mimir-10                            6.337m ±  4%   6.305m ±  3%        ~ (p=0.240 n=6)
Query/nh_100,_range_query_with_1000_steps/Mimir-10                           44.77m ±  3%   44.35m ±  6%        ~ (p=0.699 n=6)
Query/nh_2000,_instant_query/Mimir-10                                        32.90m ±  1%   32.80m ±  7%        ~ (p=0.394 n=6)
Query/nh_2000,_range_query_with_100_steps/Mimir-10                           112.8m ±  1%   114.0m ± 12%        ~ (p=0.180 n=6)
Query/nh_2000,_range_query_with_1000_steps/Mimir-10                          815.0m ±  3%   809.3m ±  1%        ~ (p=0.589 n=6)
Query/sum(nh_1),_instant_query/Mimir-10                                      141.4µ ±  2%   140.8µ ±  6%        ~ (p=0.240 n=6)
Query/sum(nh_1),_range_query_with_100_steps/Mimir-10                         220.6µ ±  7%   195.3µ ±  1%  -11.44% (p=0.002 n=6)
Query/sum(nh_1),_range_query_with_1000_steps/Mimir-10                        793.8µ ±  3%   650.6µ ±  3%  -18.04% (p=0.002 n=6)
Query/sum(nh_100),_instant_query/Mimir-10                                    1.965m ±  1%   1.943m ±  1%   -1.13% (p=0.002 n=6)
Query/sum(nh_100),_range_query_with_100_steps/Mimir-10                       6.883m ±  1%   6.813m ±  0%   -1.02% (p=0.002 n=6)
Query/sum(nh_100),_range_query_with_1000_steps/Mimir-10                      49.23m ±  1%   48.88m ±  1%        ~ (p=0.065 n=6)
Query/sum(nh_2000),_instant_query/Mimir-10                                   32.82m ±  0%   32.70m ±  1%        ~ (p=0.132 n=6)
Query/sum(nh_2000),_range_query_with_100_steps/Mimir-10                      126.7m ±  1%   126.3m ±  1%        ~ (p=0.394 n=6)
Query/sum(nh_2000),_range_query_with_1000_steps/Mimir-10                     944.9m ±  1%   929.0m ±  1%   -1.68% (p=0.004 n=6)
Query/sum_by_(l)(nh_1),_instant_query/Mimir-10                               148.7µ ±  3%   140.1µ ±  1%   -5.79% (p=0.002 n=6)
Query/sum_by_(l)(nh_1),_range_query_with_100_steps/Mimir-10                  213.8µ ±  3%   193.8µ ±  3%   -9.33% (p=0.002 n=6)
Query/sum_by_(l)(nh_1),_range_query_with_1000_steps/Mimir-10                 772.0µ ±  1%   636.1µ ±  1%  -17.60% (p=0.002 n=6)
Query/sum_by_(l)(nh_100),_instant_query/Mimir-10                             1.982m ±  0%   1.959m ±  0%   -1.16% (p=0.002 n=6)
Query/sum_by_(l)(nh_100),_range_query_with_100_steps/Mimir-10                7.617m ±  1%   6.377m ±  0%  -16.28% (p=0.002 n=6)
Query/sum_by_(l)(nh_100),_range_query_with_1000_steps/Mimir-10               53.80m ±  1%   44.88m ±  1%  -16.58% (p=0.002 n=6)
Query/sum_by_(l)(nh_2000),_instant_query/Mimir-10                            34.17m ± 18%   34.18m ±  1%        ~ (p=0.937 n=6)
Query/sum_by_(l)(nh_2000),_range_query_with_100_steps/Mimir-10               135.7m ±  2%   115.3m ±  2%  -15.02% (p=0.002 n=6)
Query/sum_by_(l)(nh_2000),_range_query_with_1000_steps/Mimir-10             1037.8m ±  3%   819.6m ±  4%  -21.03% (p=0.002 n=6)

Which issue(s) this PR fixes or relates to

Fixes #8360 (comment)

Checklist

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

[jhesketh]

Nice stuff. I think this makes sense to me. Some more tests would be good, but otherwise I think we should proceed with this approach.

[jhesketh]

lgtm, just some nits/queries

[jhesketh]

Thanks, LGTM!