feat(asap-tools): render sketch-bench e2e queries as ClickHouse SQL

[zzylol]

Builds on #837 (PromQL → ClickHouse SQL for generated2, merged); now based on main.

Summary

Renders a sketch-bench #777 workload table's e2e.queries as ClickHouse SQL, for the ClickHouse baseline of the end-to-end runs (#824). The new datasets/fake_metric/sketch_bench_workloads.py sits next to #837's generate_workloads.py and shares its extrapolatedRate (split out as rate_from_samples, whose generated text is unchanged).

• Classic templates (PromQL): parsed into a Query record, adding the shapes feat(asap-tools): run generated2 PromQL workloads on ClickHouse #837 lacks: topk(k, sum by (x) (...)), sum by (x) (rate(...)), quantile by (x) (q, instant), the ratio of two quantile_over_time, and the per-series forms.
  • Windows follow Prometheus 2.53, which the bare-metal experiments install (v2.53.2): range selectors are closed, [T−d, T], and instant vectors take each series' latest sample in [T−1s, T]. Measured on CloudLab against a real Prometheus 2.53: its [5s] sums match ClickHouse's ts >= T−5s AND ts <= T exactly, while (T−5s, T] misses the boundary samples. The paced feeder's timestamps and aligned eval times always put samples on that boundary.
  • Note for feat(asap-tools): run generated2 PromQL workloads on ClickHouse #837: generated2 still reads (T−d, T] (left-open is Prometheus 3.0). With 1 s scrapes on whole seconds, a [1m] window reads 60 samples where Prometheus 2.53 reads 61. This PR leaves it unchanged; it's your call there.
• Multi-grouping templates (structured entries, sketch-bench#209): uniqExact(value) for cardinality and quantileExactInclusive(q)(value) pooled over every sample of a group. Windows are ASAPQuery's left-open (E−W, E], since no Prometheus is involved and ASAPQuery's SQL frontend and precompute windows are (start, end].
  • Also writes <table>/asapquery.sql for the ASAPQuery arm: each entry's sql from the table as written (sketch-bench#209 renders the ASAPQuery dialect once, QUANTILE(q, col), COUNT(DISTINCT col), BETWEEN DATEADD(s, -W, NOW()) AND NOW()), with only NOW() replaced by the eval time floored to T. All 936 distinct queries of the 12 tables parse in ASAPQuery's SQL frontend to (⌊E⌋−W, W).
• Eval time: {{ end_time }} is floored to a multiple of T in the SQL, matching the aligned query client (feat(asap-tools): run sketch-bench workload tables end to end with the paced feeder #841).
• Output: one experiment_run_clickhouse.py config per table, one query group per interval_secs (repetition = T), plus init.sql with the tables (gauge/counter: ts, labels, value; records: ts, labels, value fields). Unknown shapes, capabilities or labels raise.

Testing

• Python reference: a separate implementation of the PromQL 2.53 semantics, and of the pooled multigroup semantics, over small datasets.
  • Classic: 6 series at about 4 samples/s on ms timestamps, with a late-starting counter, a series that stops and a 3 s gap.
  • Multigroup: a skewed http records table with endpoint a child of service.
• Coverage: 18 classic shapes and template 16's full CUBE plus 7 cardinality groupings, each at several eval times, all run in one clickhouse-local process. Every SQL result matches the reference within 1e-9, including empty windows and every extrapolatedRate branch.
• Mutation checks: a left-open window on classic, or a closed one on multigroup, makes 84–252 subtests fail.
• Real tables: every statement for sketch-bench's classic (108), multigroup (104) and all-metrics8 (1696) tables executes in clickhouse-local.
• ASAPQuery dialect: all 52 multigroup queries, rendered at aligned and unaligned end times, parse with ASAPQuery's SQLPatternParser/SQLPatternMatcher to (start, duration) = (⌊E⌋−W, W).
• CLICKHOUSE_BINARY=<path> python3 -m pytest -q tests/test_sketch_bench_workloads.py passes (the ClickHouse tests are skipped without a binary). The experiments suite passes. black/flake8 clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AnJ7LocWxguh3q4QWkCYae