Repository navigation
Nested rate/increase windows: Rate vs Increase accumulators, no shared PaneChain, PaneChain hides its SummaryAgg #873
Description
Workload
Three repeating PromQL queries on ae7ae6c2, all every 10 s:
- Q1
rate(data[5m]) - Q2
rate(data[1m]) - Q3
increase(data[10m])
data is a counter with 10,000 series, one sample every 15 s per series (≈667 samples/s ingesting).
Save as wl.json:
{"query_workload": {"language": "promql", "query_batch": null, "repeating_queries": [
{"query": "rate(data[5m])", "demand": {"fixed_interval": 10000},
"requirements": {"accuracy": {"max_error": 0.01, "failure_probability": 0.001}},
"predictability": {"predictable": {"known_at": null}}},
{"query": "rate(data[1m])", "demand": {"fixed_interval": 10000},
"requirements": {"accuracy": {"max_error": 0.01, "failure_probability": 0.001}},
"predictability": {"predictable": {"known_at": null}}},
{"query": "increase(data[10m])", "demand": {"fixed_interval": 10000},
"requirements": {"accuracy": {"max_error": 0.01, "failure_probability": 0.001}},
"predictability": {"predictable": {"known_at": null}}}]},
"data_workload": {"arrival": {"ingesting": {"rows_per_second": 666.67}},
"sources": {"data": {"series_count": 10000, "metric_type": "counter", "sample_interval": 15000}}}}cargo run -p asap-devtools --bin stage_pipeline -- --workload wl.json --exhaustive --out out.jsonThis gives 16 logical plans, all exact: each query either does or doesn't build its window from 10 s panes (30, 6 and 60 panes), and the scan either is or isn't shared. A query in panes runs Scan → PaneChain → SummaryMerge → FinalizeExactAccumulator.
Questions
- Rate and Increase accumulators should be the same. The rate queries build
ExactAggregate [Rate, Rate]and the increase queryExactAggregate [Increase, Increase].rateisincreasedivided by the window, so the state per series should be the same, with only the final step differing. - There should be sharing across PaneChains. The windows are nested (1m ⊂ 5m ⊂ 10m) and all use 10 s panes, so Q2's 6 panes are the newest 6 of Q1's 30, which are the newest 30 of Q3's 60. However, only the scan is shared, not the PaneChains.
- Should
PaneChaincontain an explicitSummaryAgg? TodayPaneChaincarries the build's own fields (family,input,grouping,filter,reduction) pluspane_countandpane_width_ms, and there is noSummaryAggnode beneath it. Would it be clearer to model it as aSummaryAggper pane, withPaneChainonly describing the windowing?
Activity
zzylol commentedon Oct 22, 2026
Additional matched-input control from the observability probe on
c3be284c97c1f7ecdc490d78ddecf3e5f78c2007(2026-10-22):Both queries have identical filters and grouping, repeat every 30 seconds, and require exact accuracy (
max_error: 0,failure_probability: 0). The counter has 1,000 series sampled every 15 seconds with ongoing ingestion.Exhaustive enumeration gives 8 logical candidates, and selects
P2-m1: a shared scan at query time. None of the logical candidates shares rate/increase summary state or a common pane chain. Pane variants keep separateRateandIncreasefamilies (10 and 120 panes of 30 seconds).This supports the readout/state-sharing concern in this issue independently of differing filters: the original API-server production rules also differ in filters and grouping, but this control removes those differences.
Self-contained workload (save as
matched.json, then runstage_pipeline --workload matched.json --enforce-accuracy-guarantees --exhaustive --max-combinations 10000 --max-displayed 100 --out matched.out.json):{ "query_workload": { "language": "promql", "repeating_queries": [ { "query": "sum by (cluster, code, verb) (rate(apiserver_request_total{job=\"apiserver\",verb=~\"LIST|GET\"}[5m]))", "demand": { "fixed_interval": 30000 }, "requirements": { "accuracy": { "max_error": 0.0, "failure_probability": 0.0 } }, "predictability": { "predictable": { "known_at": null } } }, { "query": "sum by (cluster, code, verb) (increase(apiserver_request_total{job=\"apiserver\",verb=~\"LIST|GET\"}[1h]))", "demand": { "fixed_interval": 30000 }, "requirements": { "accuracy": { "max_error": 0.0, "failure_probability": 0.0 } }, "predictability": { "predictable": { "known_at": null } } } ] }, "data_workload": { "arrival": { "ingesting": { "rows_per_second": 66.66666666666667 } }, "sources": { "apiserver_request_total": { "sample_interval": 15000, "series_count": 1000, "metric_type": "counter" } } } }This evidence covers planning and executor compilation, not execution equality of a proposed shared state.