MQE: skip decoding histogram buckets where possible #11587

charleskorn · 2025-06-02T05:04:42Z

What this PR does

This PR implements an optimisation inspired by prometheus/prometheus#14097.

I've implemented this as a MQE optimisation pass (rather than reusing the logic from that Prometheus PR) for two reasons:

It helps prove out the structure of optimisation passes, and demonstrate they're applicable to more than just common subexpression elimination.
We need to consider how this optimisation interacts with common subexpression elimination. Specifically, we need to consider all the different paths to a selector after common subexpression elimination has been performed to know if we can skip decoding histogram buckets. Alternatively, if we were to reuse the logic from Prometheus, the common subexpression elimination pass would need to know how to merge selectors that are identical save for the SkipHistogramBuckets hint.

Peak memory consumption is reduced by ~10% in our benchmarks. Latency improves by up to 25% in some cases, but regresses by a small amount (in both absolute and relative terms) for some instant queries:

goos: darwin
goarch: arm64
pkg: github.com/grafana/mimir/pkg/streamingpromql/benchmarks
cpu: Apple M1 Pro
                                                                                                            │ original.txt │             final.txt              │
                                                                                                            │    sec/op    │   sec/op     vs base               │
Query/histogram_count(sum(rate(nh_1[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                      244.8µ ± 4%   236.4µ ± 1%   -3.40% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10         291.7µ ± 4%   261.4µ ± 2%  -10.40% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10        791.2µ ± 1%   627.4µ ± 2%  -20.70% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                    7.719m ± 1%   7.948m ± 0%   +2.97% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10       11.85m ± 1%   10.64m ± 1%  -10.25% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10      54.10m ± 3%   41.61m ± 2%  -23.08% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                   146.2m ± 5%   150.3m ± 0%        ~ (p=0.065 n=6)
Query/histogram_count(sum(rate(nh_2000[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10      223.9m ± 3%   202.3m ± 0%   -9.63% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10    1047.5m ± 3%   789.9m ± 1%  -24.60% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                     264.6µ ± 4%   243.7µ ± 1%   -7.92% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10        359.7µ ± 4%   320.3µ ± 2%  -10.96% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10       1.295m ± 3%   1.089m ± 0%  -15.92% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                   8.048m ± 1%   8.472m ± 1%   +5.26% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10      16.53m ± 2%   15.48m ± 4%   -6.34% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10    100.81m ± 1%   84.79m ± 0%  -15.89% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                  151.2m ± 0%   160.0m ± 0%   +5.83% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10     314.8m ± 0%   295.9m ± 1%   -6.00% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10     1.968 ± 1%    1.660 ± 0%  -15.64% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                        244.1µ ± 2%   230.4µ ± 2%   -5.60% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10           289.1µ ± 2%   261.0µ ± 1%   -9.70% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10          783.2µ ± 1%   623.1µ ± 0%  -20.44% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                      7.820m ± 2%   7.937m ± 1%   +1.49% (p=0.041 n=6)
Query/histogram_sum(sum(rate(nh_100[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10         11.95m ± 1%   10.57m ± 1%  -11.52% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10        54.67m ± 1%   41.65m ± 0%  -23.83% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                     146.8m ± 1%   150.1m ± 0%   +2.27% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10        224.9m ± 0%   200.4m ± 1%  -10.90% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10      1050.1m ± 1%   790.5m ± 1%  -24.72% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                       258.7µ ± 2%   243.4µ ± 2%   -5.91% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10          353.7µ ± 3%   317.4µ ± 2%  -10.27% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10         1.290m ± 1%   1.084m ± 1%  -15.93% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                     8.085m ± 0%   8.471m ± 1%   +4.78% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10        16.58m ± 6%   15.49m ± 1%   -6.59% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10      101.78m ± 3%   84.98m ± 2%  -16.50% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                    152.6m ± 2%   160.2m ± 2%   +4.99% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10       317.7m ± 1%   296.4m ± 1%   -6.70% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10       1.980 ± 1%    1.654 ± 0%  -16.49% (p=0.002 n=6)
geomean                                                                                                        15.01m        13.55m        -9.76%

                                                                                                            │ original.txt │              final.txt              │
                                                                                                            │      B       │      B        vs base               │
Query/histogram_count(sum(rate(nh_1[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                     76.69Mi ± 5%   68.97Mi ± 5%  -10.07% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10        74.16Mi ± 2%   66.49Mi ± 1%  -10.34% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10       73.48Mi ± 1%   64.20Mi ± 2%  -12.63% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                   71.55Mi ± 1%   62.59Mi ± 3%  -12.52% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10      69.62Mi ± 1%   62.04Mi ± 2%  -10.89% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10     71.49Mi ± 1%   63.12Mi ± 1%  -11.71% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                  75.71Mi ± 2%   67.80Mi ± 2%  -10.44% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10     75.67Mi ± 2%   67.84Mi ± 2%  -10.36% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10    79.64Mi ± 1%   72.62Mi ± 1%   -8.81% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                    75.88Mi ± 2%   67.17Mi ± 2%  -11.48% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10       73.75Mi ± 2%   66.02Mi ± 4%  -10.49% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_1[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10      73.00Mi ± 1%   64.12Mi ± 2%  -12.16% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                  70.35Mi ± 2%   63.04Mi ± 2%  -10.39% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10     69.81Mi ± 1%   61.73Mi ± 1%  -11.57% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_100[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10    71.08Mi ± 1%   62.70Mi ± 1%  -11.78% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                 74.77Mi ± 3%   67.80Mi ± 3%   -9.32% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10    75.47Mi ± 1%   66.90Mi ± 2%  -11.36% (p=0.002 n=6)
Query/histogram_count(sum(rate(nh_2000[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10   79.35Mi ± 1%   70.52Mi ± 2%  -11.14% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                       76.41Mi ± 2%   69.45Mi ± 4%   -9.10% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10          75.35Mi ± 3%   66.16Mi ± 2%  -12.19% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10         72.92Mi ± 1%   63.72Mi ± 1%  -12.62% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                     71.88Mi ± 2%   62.25Mi ± 1%  -13.40% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10        70.02Mi ± 1%   61.74Mi ± 1%  -11.83% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10       71.70Mi ± 1%   63.26Mi ± 1%  -11.78% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[2m]))),_instant_query/engine=MimirWithQueryPlanner-10                    74.36Mi ± 2%   66.23Mi ± 2%  -10.94% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[2m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10       75.64Mi ± 1%   67.31Mi ± 1%  -11.01% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[2m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10      79.44Mi ± 1%   72.38Mi ± 1%   -8.89% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                      75.46Mi ± 4%   67.14Mi ± 2%  -11.03% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10         73.41Mi ± 3%   65.93Mi ± 3%  -10.19% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_1[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10        72.95Mi ± 1%   63.78Mi ± 1%  -12.57% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                    71.73Mi ± 2%   63.17Mi ± 2%  -11.93% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10       69.86Mi ± 1%   61.66Mi ± 1%  -11.73% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_100[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10      71.34Mi ± 1%   62.35Mi ± 2%  -12.60% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[20m]))),_instant_query/engine=MimirWithQueryPlanner-10                   75.13Mi ± 1%   67.05Mi ± 3%  -10.75% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[20m]))),_range_query_with_100_steps/engine=MimirWithQueryPlanner-10      75.03Mi ± 1%   67.27Mi ± 2%  -10.35% (p=0.002 n=6)
Query/histogram_sum(sum(rate(nh_2000[20m]))),_range_query_with_1000_steps/engine=MimirWithQueryPlanner-10     78.70Mi ± 2%   70.84Mi ± 1%   -9.98% (p=0.002 n=6)
geomean                                                                                                       73.92Mi        65.69Mi       -11.13%

⚠️ This PR is pending prometheus/prometheus#16686 being merged and vendored into Mimir.

Which issue(s) this PR fixes or relates to

#10067

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.

github-actions · 2025-06-02T05:06:30Z

💻 Deploy preview available:

charleskorn · 2025-06-04T04:27:03Z

This PR is pending prometheus/prometheus#16686 being merged and vendored into Mimir.

charleskorn added 6 commits May 19, 2025 12:39

Add hint field and pass through to operator

8ea38f6

Initial implementation

8748782

Try to reuse iterator where possible

9207e7d

Add config flag

7925012

Move utility functions useful to all optimisation pass tests

358707f

Initial implementation of optimisation pass

0369ebb

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

MQE: skip decoding histogram buckets where possible #11587

MQE: skip decoding histogram buckets where possible #11587

Uh oh!

charleskorn commented Jun 2, 2025 •

edited

Loading

Uh oh!

github-actions bot commented Jun 2, 2025

Uh oh!

charleskorn commented Jun 4, 2025

Uh oh!

Uh oh!

MQE: skip decoding histogram buckets where possible #11587

Are you sure you want to change the base?

MQE: skip decoding histogram buckets where possible #11587

Uh oh!

Conversation

charleskorn commented Jun 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does

Which issue(s) this PR fixes or relates to

Checklist

Uh oh!

github-actions bot commented Jun 2, 2025

Uh oh!

charleskorn commented Jun 4, 2025

Uh oh!

Uh oh!

charleskorn commented Jun 2, 2025 •

edited

Loading