[Refactor] Pane-width rounds without cross-stage calls: rerun Pass 2 per round, or pass an incremental Pass 2 state #1048

Example

Workload tools/dag-viewer/memo-examples/promql/ratio-of-quantiles-5m.json: one query, evaluated every minute, with max_error 0.01:

avg by (job) (quantile_over_time(0.5, data[5m]) / quantile_over_time(0.5, data[5m]) + quantile_over_time(0.9, data[5m]))

Over the complete pane-width domain, Pass 2 gives this one query 336 PaneChain shared-build groups (production numbers from #1010). That is too many to generate before a first plan exists. So the search (candidate_connection_search in crates/plan-selection/src/search/route.rs) runs in rounds:

1. Seed round. PaneWidthDomain::SeedGrids, budget min(¼ of the remaining time, 5 s).
2. Wider rounds. PaneWidthFrontier::next_domain yields Expanding { max_parts } domains, then Complete. Each non-complete round gets at most 5 s and 16 evaluations.
3. Merge. The best validated plan across rounds is kept (merge_validated_rounds).

Between rounds, Pass 2 is incremental:

• DomainProgress.local_memo keeps the memo from before state-group expansion.
• extend_local_window_composition(&mut memo, entries, previous_domain, domain, deadline) adds only the widths new in this round.
• expand_concrete_summary_state_groups_before then runs on a clone, because expansion can run only once on a memo.

Rounds run only when the workload has a pane window, the search is not exhaustive and does not keep every candidate, and the budget allows it. stage_pipeline with keep_every_candidate therefore always runs one Complete round.

The decision

#1047 moves the rounds into the asap-planner facade (Pass 2 → Stage 2 → Stage 3 per round) and forbids calls across stages. Pass 2 then needs one of:

• A. Stateless (proposed first step). generate_logical_asap_memo_candidates(&pass1, entries, pane_widths, deadline) regenerates Pass 2 from scratch each round. This is simple, and no hidden state crosses stages. It repeats the window composition for widths earlier rounds already produced.
• B. Incremental handle. Pass 2 returns its output plus an opaque Pass2State (today's local_memo + previous_domain). The facade passes the state back into the next round without reading it. This saves the repeated generation, but it adds a type that lives across rounds.

Proposal: implement A, measure it, and switch to B only if the measurement requires it.

Acceptance

Measure on the same host with a release stage_pipeline, with rounds enabled (no --exhaustive, default keep_every_candidate = false), on:

• promql/ratio-of-quantiles-5m.json
• the five promql/rangeheavy-*.json
• the random mixed workloads

Compare A against main:

• the selected plan and its cost
• the number of rounds completed within the budget
• Pass 2 time per round
• total planning time
• peak RSS

A is kept if the selected plans are identical and the total planning time and peak RSS are within the #1047 threshold (10%). Otherwise implement B and measure again.

Related

• Parent: [Refactor] One named type per stage boundary: PerComputationLogicalMemoCandidates → LogicalASAPMemoCandidates → PhysicalASAPMemoCandidates #1047 (stage-boundary types; no calls across stages)
• Reorganize by stage: one memo-based Pass 2, Stage 1 finishes the memo, deployment inputs split by stage, asap-planner as the only entry point #958 (reorganize by stage)
• Stage 1 Pass 2 review: ASAP-aware CSE rules over memo groups and summary-state groups #1013 (Pass 2 review)
• Tracking: plan every whole evaluation workload within 3 minutes (exact pruning, memo-group search) #879 (plan whole workloads within minutes)