feat(asap-tools): exact materialized-view arm for sketch-bench e2e queries

[zzylol]

Stacked on the sketch-bench SQL PR (base branch feat/e2e-sql-from-table).

Summary

Adds the baseline_mv arm: ClickHouse with materialized views and no sketches, only exact, mergeable states. Each view's bucket width is the query's repeat interval T.

• One view per (table, value column, statistic, T), an AggregatingMergeTree keyed by bucket and the union of the groupings that read it, plus their child_of parents. Per-series and instant queries keep every label. Coarser groupings GROUP BY a subset and merge (roll-up). For example, template 16's whole CUBE reads one (region, service, endpoint) view.

• States:

  Statistic	State
  sum, count	sumState, countState
  quantile	quantilesExactInclusiveState over the view's levels
  distinct count	uniqExactState
  instant vector	argMaxState(value, ts)
  rate/increase	per series: first/last value, min/max ts and count, fed into the same extrapolatedRate
  top-k	exact per-key sums, then ORDER BY … LIMIT k

• Windows: bucket b holds (b−T, b], and an edge flag keeps samples at exactly b apart. Classic's closed [E−W, E] merges the buckets in (E−W, E] plus bucket E−W's edge rows. Multigroup's left-open (E−W, E] merges just the buckets. A window that isn't a multiple of T raises. Instant queries need T = 1 s.

• Per-arm init SQL, so the baseline arm never pays for the views: init.sql (tables only) and init_mv.sql (tables, then views; the views must exist before data arrives). Both drop old views first.

  • experiment_run_clickhouse.py now uses the run's first mode for ingest monitoring when baseline isn't in the run, so a baseline_mv-only run works.

Testing

• On the same datasets as the SQL PR, every MV query matches its raw SQL in clickhouse-local over 392 eval points, classic and multigroup. Distinct counts, quantiles, per-series stats and top-k keys match bit-exactly; sums match within 1e-9 (the summation order differs).
• Mutation checks: ignoring the edge rows, or using left-closed toStartOfInterval buckets, makes 84–132 subtests fail.
• Real tables: every MV statement and init_mv.sql execute for classic (4 views), multigroup (3 views: http_user_id_cardinality_60s, http_latency_quantile_60s, flows_src_ip_cardinality_60s) and all-metrics8 (56 views).
• New runner test in tests/test_experiment_run_clickhouse.py. black/flake8 clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AnJ7LocWxguh3q4QWkCYae

[milindsrivastava1997]

Code review

Stacks on #844. Core MV logic checks out (closed and left-open windows, -State/-Merge pairs, child_of rollups), and test_mv_sql_matches_raw_sql is good coverage.

1. Medium — sketch_bench_workloads.py ~L535 (init_sql): only views from the current table file are dropped. If workload A's baseline_mv run created data_value_quantile_60s, workload B on the same data table (different stats/intervals) drops and recreates data but never drops A's view, which likely keeps firing on inserts (untested) and skews B's insert cost. Drop all views sourcing the metric tables (via system.tables dependencies) or use a per-stem prefix.
2. Medium-low — ~L584 (convert): when the mv arm is skipped (unservable query), stale init_mv.sql, interval_<T>s_baseline_mv.sql, <stem>_clickhouse_mv.yaml from an earlier run remain; a sweep over *_clickhouse*.yaml runs them with misaligned query indices. Delete them on skip; test_unservable_query_skips_mv_arm uses a fresh dir so can't catch this.
3. Low — ~L425 (plan_views): name = parent[name] raises a bare KeyError if child_of names a label not in the metric — inputs that converted fine before now crash with no context. Validate and raise a ValueError naming table and label.
4. Low — ~L577 vs ~L413: "window must be a multiple of interval" is checked in both convert and plan_views; a new unservable shape added to only one makes convert crash instead of skip. Share one mv_servable(q) predicate.
5. Low — ~L443-510: view_states, mv_agg, raw_agg each branch on the same stat names; a per-stat table (state cols, -State, -Merge) would keep them in sync.
6. Nit — docstring ~L29-31: uses T where it means eval time E (E - W, [E - 1s, E]).

Tests: missing a regression test that baseline init.sql drops existing views, and one for re-converting into a dir that already has mv files.