Repository navigation
One workload-level PhysicalASAPDAG and PlanOutput deduplication — after #511 #521
Description
Goal
#509 defines a candidate as a DAG for the whole workload, and selection returns one PhysicalASAPDAG with each shared summary costed once. Today:
PlanOutputholds oneQueryLifecyclePlanper query.PhysicalASAPDAG(renamed in refactor!: rename planner candidate types to the #509 stage names #514) is compiled from one root. A summary shared by feat: share identical summary producers across queries after Pass 1 #515 appears in several plans, and consumers dedupe it byRcpointer.- For a root that reaches a shared state, assembly's raw-vs-summary check (feat: share identical summary producers across queries after Pass 1 #515) compares that root's raw cost over the union's reads with the whole shared state. This is approximate unless the members cost about the same per read.
- Planning and selection of a candidate run per target group rather than as one whole-workload candidate costed as a unit.
CandidateDagInventory/enumerate_candidate_dagsexist, but only tests use them.
Blocked on
Implementation of #511 (unified operator graph: OperatorNode with common metadata, no KeepPreAsap wrappers). Do not start before #511 is implemented.
The merged per-workload graph should be built on #511's common operator graph rather than on today's per-root post-ASAP DAG.
Scope
- Assemble the selected per-root plans into one workload DAG with shared nodes deduplicated, and compile that DAG into one
PhysicalASAPDAG. PlanOutputcarries the workload-level plan, each query's root ID, and deduplicated deployments.- Replace feat: share identical summary producers across queries after Pass 1 #515's approximate shared-root raw-vs-summary check with a single whole-workload comparison.
- Decide whether selection costs complete whole-workload candidates directly, or keeps per-group selection with class-level sharing (feat: share identical summary producers across queries after Pass 1 #515).
Activity
zzylol commentedon Oct 15, 2026
Closing as done. On main the workload is planned, costed and exported as one DAG.
PhysicalASAPDAG, with one memo for the whole workload (crates/asap-aware-mapping/src/stage2_physical_asap_optimization.rs:optimize_physical_asap→compile_physical_asap_workload_with_node_ids, feat(planner): physical ASAP-aware optimization and plan selection for candidate logical ASAP DAGs #576).PlanOutput::execution_timed_dag(crates/asap-aware-mapping/src/pass/mod.rs, feat(planner): export a planned workload as one physical ASAP DAG #666) exports the selected plan the same way. Operator queries becomerootsand scalar queries becomescalar_roots, and each shared sub-DAG is exported once (crates/types/src/ir/physical_export.rs:compile_physical_asap_workload).stage1_to_stage3.rs:stage1_logical_asap_optimization, feat(planner): compose candidate logical ASAP DAGs from Pass 1 per-computation candidates #575). Stage 3 prices each candidate's single physical DAG node by node, so a shared node is costed once (crates/asap-aware-mapping/src/stage3_plan_selection.rs:select_plan→price(&candidate.dag, …), feat(planner): physical ASAP-aware optimization and plan selection for candidate logical ASAP DAGs #576). This replaces the per-group selection and feat: share identical summary producers across queries after Pass 1 #515's approximate shared-root raw-vs-summary check. feat(planner): run the planning stages as the planner pipeline; retire MajorPass #674 retiredMajorPass.One leftover:
PlanOutput.plansstill lists one root per query over sharedRcnodes. Its doc comment still says consumers dedupe byRc::ptr_eq.execution_timed_dag()andoperators()already give the deduplicated view.