Repository navigation
test(devtools): plan snapshots for 33 small single-purpose workloads (#1278) - #1284
milindsrivastava1997 wants to merge 35 commits into
Conversation
…nes they replace (#1207 PR 0)
…iming, lifetime }
Deletes Schedule and the σ-only InvalidPhysicalProperties variants; the only
invalid combination left is (IngestionTime, OneEvaluation). satisfies_physical
checks timing x lifetime only; delivers_on is the per-context alignment check
(w | I and w | φ). A OneEvaluation result runs in one evaluation context, so
x_{κ,v} covers it as for ℓ = None.
…physical drops memo InvalidLogicalProperties gains SketchBuildWithoutParameters and renames FloorNotInParameterList to ParametersNotInList, returned by LogicalProperties::new. LogicalRequirement::new no longer checks list membership (a floor is any value of the family's parameter type, e.g. heap_size >= k), so it returns Self. satisfies_physical takes only the producer and consumer.
Context::QueryEvaluation(QueryId) becomes Context::Evaluation(EvaluationSchedule):
queries with the same (interval, phase) are evaluated together and share
their OneEvaluation results. evaluation_schedule returns that schedule.
QueryId stays for z_{q,v}. A non-periodic query has no context yet.
InvalidLogicalProperties -> LogicalPropertiesError, InvalidPhysicalProperties -> PhysicalPropertiesError, LogicalRequirement -> LogicalRequirements (parameter_floor -> parameter_floor_for_required_accuracy), satisfies_logical / satisfies_physical -> producer_satisfies_logical_requirements / producer_satisfies_physical_requirements, Medium -> StorageMedium, Context / ContextId -> ExecutionContext / ExecutionContextId (x_in_context / y_in_context -> x_in_execution_context / y_in_execution_context). PhysicalCandidate and PhysicalCandidateKey fields -> logical_candidate_id, physical_properties; edge fields -> producer_candidate_id, consumer_candidate_id; PhysicalASAPMemoCandidates fields -> logical_candidates, physical_candidates.
Deletes EvaluationSchedule and ExecutionContext::evaluation_schedule. An
evaluation context is Evaluation { recurrence, range_evaluation }: queries
with equal ones share it and its OneEvaluation results, one-off queries
included. ExecutionContext, ExecutionContextId and RefreshWindowId move to
asap_types::stages so PhysicalCandidate::delivers_on can take the context.
Contexts are deduplicated by equality (QueryRecurrence holds an f64).
PhysicalASAPDAGValidationError::InvalidPhysicalProperties becomes
PhysicalProperties { node, reason }. In the files this unit edits, "kept"
describing a materialized result or retained data becomes "materialized"
or "retained" (cost, materialized views, realizer, installable plan,
deployment capabilities; locals kept/raw_kept_rows -> retained/
raw_retained_rows), and "reader" becomes "query" where it is a query
(execution demand, ScheduleGroup::queries, cost usage, labels, tests) or
"consumer" where it is a consuming candidate.
find_or_add_group returns the node rules' SchemaDerivationError and add_candidate a CandidateError (the node rules, or an output that differs from the group's). Bodies stay todo!() until PR 2.
…mo (#1207 PR 0) Memo::groups is private, read through groups(); groups are added only through find_or_add_group and queries through add_query (todo!() until PR 2). memo_of_queries returns Result and its body is todo!() until PR 2, which builds Stage 0 on find_or_add_group; Stage 0's node-level CSE (cse.rs), its test_support export and the dead share_identical helper are deleted.
…OperatorInput; coverage only in the memo (#1207 PR 0) MemoGroup { expression, group_properties: GroupProperties, parents, queries, candidates } replaces logical_node and child_groups; GroupProperties { schema, result_kind, summary_coverage, shareable_when_equal } is derived once by derive_group_properties, which applies the existing rules to the expression over MemoGroupAsInputs. The rules (output_schema, output_kind, validate_inputs, is_shareable_when_equal, summary builds, window chains, scalar subqueries, top-k, populations) are generic over OperatorInput, implemented for Rc<OperatorNode> and MemoGroupAsInput; only how they read an input changed. BuiltOrInput carries the builds a rule makes while deriving (one build over a chain's window, the build a rollup or a selection stands for). SummaryCoverage { definition: SummaryStateDefinition { input_group, unlifted_operators }, selection } is derived only from memo groups (derive_from_inputs) and stored in GroupProperties; node-tree coverage, NodeToSummaryCoverageCache, derive_cached and OperatorNode::coverage are deleted, so node trees have no merge check. Memo::logical_node_trees is deleted: Pass 1 target discovery (PR 3), devtools group labels (PR 9) and the export's view coverage (PR 6) are todo!(). find_or_add_group and add_candidate derive the properties, then todo!() until PR 2.
…elpers (#1207 PR 0) One state's θ is no longer sized for several consumers: θ comes from parameter_list per candidate. size_for_all_readers had no non-test caller; min_recorded, kll::shared_k and univmon::shared_heap served only it.
…perators; delete merged_window_build_node (#1207 PR 0) - BuiltOrInput / BuiltInput renamed InputOrTemporaryBuild / TemporaryBuild: a build a rule constructs only while it derives, never stored in the memo. - SummaryStateDefinition::unlifted_operators is Vec<Operator<Option<MemoGroupId>>>: None marks an operator's main input, Some(g) another input group it reads (a scalar subquery's). States whose definitions read subqueries compare and merge; CoverageError::SubqueryInDefinition deleted. - merged_window_build_node deleted with its test; the accuracy guarantee of a read of merged tumbling windows is todo!() until PR 4 rewrites it over groups.
…, TimeBound, Pass 2 pruned list, Stage 2 entry (#1207 PR 0) - Delete WindowLayout, LogicalProperties::window_layout and its two error variants, SummaryMerge::newest_tumbling_windows and SummaryTumblingWindowChain::tumbling_window_count, with the count-based chain builds and expansion; behaviour that read the count is todo!() for PR 1, 2, 6, 7 or 9. - SelectionBox { columns } only, and TimeBound { At, FromEvaluationTime }; lifting TimeRange/TimeShift is todo!() for PR 2. - PrunedLogicalCandidate, PruneCause, RootPruneReason and LogicalASAPMemoCandidates::pruned. - error_propagation(consumer, metric) -> Option<ErrorPropagation>; Memo::set_input_requirements. - PhysicalCandidateKey deleted (PrunedCandidate<PhysicalCandidate>), PrunedCandidate::group and PhysicalASAPMemoCandidates::query_roots dropped; new(logical, physical, edges, pruned). - Stage 2 entry takes &DeploymentCapabilities (moved into asap-types) and cannot fail; PruneReason follows the doc's Stage 2 table. - DataArrival::{Ingesting, Mixed} gain max_out_of_order_ms. - GroupProperties::summary_coverage -> aggregate_coverage.
…l; ScalarValue::Timestamp (#1207 PR 0) Replace TimeBound by IntervalConstraintLimit { FixedValue(ScalarValue), EvaluationTimePlusMs(i64) }: Constraint::Interval holds Bound<IntervalConstraintLimit> on every column, so the timestamp column and value columns share one interval type. Doc comments carry the examples (m[5m], offset 1h, @ T, SQL ts >= literal, price > 10). ScalarValue::Timestamp(ms since the epoch) is the fixed value of an @ T or a SQL timestamp literal.
…s; delete ParametersOfAnotherFamily (#1207 PR 0) A range m[r] evaluated at t reads (t - r, t]: IntervalConstraintLimit and Constraint::Interval examples, the chain window doc and the primitive-schema table now use Excluded lower and Included upper ends. A parameter list holds only its own family's parameters, so ParametersNotInList also rejects another family's parameters.
…acy, the plan is one PhysicalASAPDAG (#1207 PR 0) - Terminology: ops' `family` is `summary_type`; FieldDataType::Sketch(SketchAlgorithm) (SketchKind, GroupingStrategy, HydraKind and the ops' `grouping` deleted); FamilyCapability is SummaryTypeCapability; parameter_list(algorithm). - θ is structure only: SketchParams derives equality; KLL failure_probability, UnivMon distinct_failure_probability and HLL/KMV relative_error deleted. - ValueSign and GroupProperties::column_sign; SourceFacts::column_signs replaces values_non_negative. - LogicalRequirements { required_accuracy: Option<RequiredAccuracy> }; RootPruneReason::BelowRequiredAccuracy; Memo::add_query takes the target and n_q; AggIntent measures carry no accuracy and front ends take none. - asap-accuracy: AccuracyEvidence::measured_guarantee, theoretical_guarantee (was local_guarantee), producer_satisfies_logical_requirements(memo, producer, consumer, input_index, evidence); the AccuracyModel trait and evidence provider, Pass 1's accuracy pruning and AccuracyCheckInputs deleted. - Physical: ExecutionContext::Evaluation { grid: EvaluationGrid }, IngestionPipeline(RefreshWindow); RefreshWindowId deleted; producer_satisfies_physical_requirements(producer, consumer_timing); PruneReason::CountMinWeightsNotProvenNonNegative. - Plan: PlanOutput { plan: PhysicalASAPDAG }; PhysicalASAPDAGNode:: selected_physical_candidate; select_plan returns (PhysicalASAPDAG, Selection); Selection { accepted_accuracy_per_query, pruned_candidates, milp_* }; RealizedPlan, PhysicalWorkloadPlan, DagOfSelectedCandidates, SelectedEdge, SelectedNodeId, AssignmentRealizer, ScanWindow and the scan/view window functions, SearchReport, SearchStatus, InexactObjective, ObjectiveDiffersFromPlanCost, SelectedPlanFailedCheck and PlanCost::cost_unit deleted. - Plan file wire version 14 (summary_type, Sketch(algorithm), no grouping). - Cost names: NodeCostSource, cost_source, PlanRanking::Cost; Serialize on PruneReason, PhysicalProperties, OriginOfCandidateForMemoGroup and RootPruneReason.
…e trait, MemoQuery, evaluation grids of groups (#1207 PR 0) 2026-11-09 decisions: - theoretical_guarantee and AccuracyEvidence::measured_guarantee take the failure probability and return the error bound that holds at it (None when θ cannot fail that rarely); bodies stay todo!() for PR 4. - AccuracyEvidence is a trait; NoMeasuredGuarantees has no measurement. PlanningModels::evidence and Pass 2 take &dyn AccuracyEvidence. - The memo holds its queries in one Vec<MemoQuery> { root, accuracy_target, values_per_evaluation }; add_query(query). - SupportedSummary::family is summary_type. - satisfies is deleted with metric_meets, scope_meets, native_target and AccuracyScope::covers, which only it used; target_limit is the comparison limit with the tolerance. - PhysicalCandidate::delivers_on(&PhysicalASAPMemoCandidates, context). - GroupProperties::depends_on_evaluation_time (derived in PR 2) and LogicalASAPMemoCandidates::evaluation_grid_of_group (filled in PR 4).
… the tests that use it (#1207 PR 0) pass1/composition.rs composed one whole-workload candidate DAG from one per-computation candidate per target, for tests only; planning adds every way to compute to the memo instead. It and the 74 tests and fixture helpers that compose through it are deleted (no interim code for PR 3).
…racy argument (#1207 PR 0)
Move-only. Split asap-logical-asap-optimization into asap-stage0-memo, asap-stage1-pass1-per-computation-candidates and asap-stage1-pass2-reuse-candidates; rename asap-physical-asap-optimization to asap-stage2-physical-optimization and asap-plan-selection to asap-stage3-plan-selection. The old test_support splits by caller. The crate-graph test checks the new crates; no stage crate depends on another.
Pass 1's tests that build Stage 0's memo and Stage 0's tests that lower PromQL through Pass 1 are deleted, with the dev-dependencies and helpers only they used. lower_promql moves to Pass 1's test fixtures, so Pass 1 no longer depends on the PromQL frontend outside its own tests. The crate-graph test now also checks dev-dependencies between stage crates.
…1207 PR 0b) Frontend tests that compose Pass 1 candidates and Pass 1 tests that lower PromQL through the frontend are deleted, with the helpers, dev-dependencies and test targets only they used. MetricsQL's test that compares with PromQL's lowering goes too, so no frontend tests depend on another frontend. The crate-graph test checks dev-dependencies between stage crates with no exceptions.
…R 0b) The PromQL test that planned through asap-planner is deleted with its planner_support helpers; the PromQL and SQL frontends drop their asap-planner dev-dependency, and PromQL its unused asap-accuracy one. The crate-graph test now also forbids a stage crate's dev-dependency on the facade.
- LogicalProperties::new(expression, parameters): a sketch build (a chain included) needs θ from parameter_list(algorithm) of its FieldDataType::Sketch(algorithm); an exact operator has none. - parameter_list(algorithm): structures only. Count-Min lists each width with depths 5, 7, 10 and 12, Count Sketch with 47, 81, 117 and 155 (its median-of-rows δ levels 1e-2 … 1e-5); heap variants each of those with the 1-2-5 heap sizes; UnivMon 40 entries; KLL, DDSketch, HLL, KMV one entry per ε level. A test keeps the lists equal to today's sizing, which is test-only. - PhysicalProperties::new rejects (IngestionTime, OneEvaluation); producer_satisfies_physical_requirements(producer, consumer_timing). - PhysicalCandidate::delivers_on(candidates, context): None in every context, OneEvaluation only in Evaluation contexts, a chain with AcrossEvaluations when its width divides the grid's period and phase (or the refresh window's width and phase), any other AcrossEvaluations result everywhere when its group does not depend on the evaluation time, else when its group's evaluation grid holds the context's times; Materialized(v) contexts stay todo!() for PR 7. - Chains: one schema (window_end at column 0, keys, state), TimeRange and TimeShift over a chain select windows (an instant TimeRange is rejected), a SummaryMerge over them finalizes and rolls up as the chain's build over that range (the two PR 0 todos). - summary_type_of_operator; QueryRecurrence::evaluation_grid_ms is None for a fixed interval without a phase.
…, depends_on_evaluation_time, aggregate_coverage), @ end()/@ start() lifted above their subtree (#1207 PR 2) - Memo: hash-consing (find_or_add_group by expression hash), candidate storage (add_candidate), add_query(MemoQuery); Memo::new(declared column signs), private declared_column_signs read by a Scan's column_sign. - Stage 0: memo_of_queries(roots, accuracy_targets, data) adds each query with its accuracy target and n_q (values_per_result over series_read), and collects the declared signs of the scanned sources (a counter's value is non-negative without a declaration). - GroupProperties::column_sign derived per operator (concepts doc, "Value sign"); depends_on_evaluation_time: a TimeRange or time()/now() makes a result depend on the evaluation time, @ T pins it; reads_only_time_series: a Scan of a time series, or every input group reading only time series (OperatorInput::reads_only_time_series computes the same rule over node trees; the recursive walks in topk.rs and Pass 1 are deleted). - aggregate_coverage for summary states, chains, TimeRange/TimeShift over a chain, merges and exact Aggregates; timestamp constraints are IntervalConstraintLimit intervals (t - o - W, t - o]; SQL timestamp literals and identity casts fold so absolute filters lift. - PromQL: @ end() / @ start() are lowered as TimeShift { at } above the selector or subquery they anchor, which is lowered without them, so it is the group of the same expression without @; coverage stops at that TimeShift (it is the definition's input group).
…1207 PR 2b) GroupProperties gains column_lineage (one ColumnLineage per output column, derived once from the input groups'). Lineage and column_sign are derived together by one per-operator column layout (stages/column_properties.rs, which replaces column_sign.rs): a copied column keeps its lineage and sign, a computed one is Computed with the operator's sign rule. Chains compute window_end and their state and keep their keys' lineage; a TimeRange or TimeShift over a chain, and TimeShift { at: End | Start } above a selector, pass lineage through; a merge of a chain's selected windows has the chain build's layout. Summary coverage names a selection's column by its SourceColumn lineage (the one SourceColumn type, beside Source); ColumnIdentity is deleted. A predicate on a Computed column, or on one of two columns of one lineage, is not lifted; the regrouping key check compares lineage. Schema fields and the IR Project lose their table qualifier; the front ends bind qualified names with their own scope while lowering. The requalification patches (with_aggregate_qualifiers, node.rs) are deleted.
…1207 PR 3) Pass 1 fills Stage 0's memo with each group's initial candidates, every intermediate result its own group and every candidate one operator over input groups: - computation_targets walks memo.groups(): one target per measure of each Aggregate group with a realization besides plain exact execution. - add_pass1_candidates adds every group's own expression, exact accumulators (state group + FinalizeExactAccumulator; avg from sum and count), one SummaryAgg state group per sketch algorithm with one build candidate per configuration of parameter_list(algorithm) and separate SummaryEstimate readout groups, fused builds (whole-expression top-k heaps; nested sum/min/max merges) listing the replaced groups in fused_groups, multi-measure joins, and key-only HAVING pushed below in the consumer Filter group's candidates. - Pass 1 sets no input requirement and no accuracy floor: every candidate's LogicalRequirements are empty; Pass 2 gives each readout its required accuracy (a top-k readout's heap_size >= k included). - ComputationTarget and FusedComputation move into Pass 1 (crate-private), FusedComputation::absorbed_groups replaces target indices, HavingEntry, HavingPlacement and Pass1Error are deleted; pass1_memo returns PerComputationCandidateError. - types: topk::ranked_column_of_group (ranked_column becomes pub(crate)); one unique-key rule for a top-k's ranked rows, exact or a heap readout; any_measure_filtered generic over the predicate's input. - Tests build their memo through the Memo methods; the integration tests stage1_logical_asap_optimization and per_computation_candidates_sql run SQL and PromQL queries through Stage 0 and Pass 1.
…requirements, dead-candidate elimination, chain groups (#1207 PR 4) Pass 2 runs, in order: the general rollup rule with one synthesized finest common group per cluster of equal definitions (replaces scan widening, finer-aggregate reuse and the rollup rules); top-k reuse; reach, caps and readout requirements (one readout candidate per requirement level, written with Memo::set_input_requirements); dead-candidate elimination with every root cause (PrunedLogicalCandidate); chain groups SummaryTumblingWindowChain[w](X) for the surviving state groups, widths from the gcd closure of the reaching queries' grids, the window and the offsets, then elimination again; and the logical edges between survivors. A merge, and a Filter/TimeRange/TimeShift over states, carries the θ of the states it reads (one candidate per θ, reading only producers with that θ). The metric rule cuts edges from readouts whose metric the consumer's input does not accept (accepted metrics pushed down from the targeted queries), not candidates; a candidate cut off from an input is eliminated too. A whole-result δ with n_q unknown prunes the readouts it reaches (ValuesPerEvaluationUnknown). The summary-input and window-rollup rules and the local-window-composition step are deleted. - types: LogicalRequirements::error_propagation(memo, consumer, metric) with the constant contracts (key-only Filter by column lineage, abs / clamp / negation, min / max / avg / count, + - / *), b = 0.1 private; AccuracyMetric::ALL; RootPruneReason::{HeapSmallerThanTopK, ValuesPerEvaluationUnknown}; MetricRule.accepted_metrics; PruneCause.through_pruned_candidates; Memo::producers; add_candidate checks θ against the group's state (CandidateError::LogicalProperties); a candidate's input requirements are part of its identity in add_candidate; summary coverage accessors no longer test-only; origins SummaryInput and WidenedScans deleted. - accuracy: theoretical_guarantee at a failure probability (KLL and KMV evaluated at it, others checked against their own), producer_satisfies_logical_requirements returns the failing root reason, readout_metric; node-tree composition, propagation, budget split, function rules, PropagationStats, test_support and the test-only sizing deleted. - devtools: one pass2 stage.
…natives; Count-Min weights via column_sign (#1207 PR 5) Stage 2 expands each logical candidate Pass 2 kept into its offered (timing, lifetime): ingestion time for time-independent groups and their inputs (counter states only over ordered rows), AcrossEvaluations for chains and grid-evaluated groups. Prune reasons: NoIngestionPipeline, RawDataNotRetained, QueryTimeResultsNotRetained, CountMinWeightsNotProvenNonNegative, InputHasNoBuildableProducer, ReachesNoQueryOutput. Physical edges carry alpha^log and alpha^phys; PhysicalASAPMemoCandidates indexes the admissible ones. Deletes WeightDomain, NonNegativeWeightProof and SummaryUpdate::weight_domain (the state group's column_sign replaces them), Pass 1's copy of reads_only_time_series, and per-crate memo test fixtures (asap_types::test_support).
…ive tumbling-window chains and a time-indexed store in the executor (#1207 PR 6) Stage 3 costs each physical candidate (w_v) and each selectable edge (c_{u,c,i}) with asap-cost, solves the baseline MILP over x, y, z, t (constraints (A)-(F), (E) only inside strongly connected components, a query-time result that is not materialized read by at most one consumer), and builds the PhysicalASAPDAG of the optimum directly, each node with its timing, lifetime, θ, retention and, materialized, its summary coverage. asap-cost sizes summaries from the node's θ and prices chains and raw data by the node's retention_ms. The executor builds and runs summaries with θ, runs tumbling-window chains natively (TimeRange/TimeShift over a chain select windows), keeps materialized rows in a time-indexed store with retention, and evaluates TimeShift{at} once per request. Installation no longer expands chains.
…}, h_v and constraint (R); asap-cost computes once per context (#1207 PR 7)
…tioned DDSketch guarantee, sum/avg relative contracts, HLL distinct count from the source's distinct bound (#1207 PR 8)
… by layer and every pruned candidate with its cause; planner_fuzz costs witnesses over Stage 3's plan; viewer reads the plan; snapshots and pins re-baselined (#1207 PR 9) - asap_planner text and asap-stage-pipeline/v2: Pass 2 logical edges with alpha^log and why each fails, Pass 2 eliminated candidates with root-cause traces, Stage 2 physical edges with alpha^phys, Stage 2 and Stage 3 pruned candidates with reasons, Stage 3 plan and plan_cost; the text folds a group's candidates that differ only in theta into one line (indices, theta values, pruned counts per cause), --full prints everything; L1/P1 lists, memo node table and same_candidate deleted. - planner_fuzz: plans with accuracy not enforced; witnesses extend Stage 3's PhysicalASAPDAG and are costed by plan_cost; shared nodes are materialized for one evaluation only within one evaluation grid; LongerWindow and NewestTumblingWindows are a TimeRange over the same chain; the accuracy relations (a-looser-accuracy, tighter-accuracy) are deleted. - PruneReason, RootPruneReason and OriginOfCandidateForMemoGroup serialize snake_case; RejectionCause keeps only Infeasible; ErrorSelectingPlan::Cost deleted. - price -> cost wording; develop docs and planner-layering refreshed; viewer reads the new document. - Every in-repo workload and test fixture declares its phase (fixed_interval_at, phase 0). - Corpus and sketch plan snapshots re-baselined; obsolete tests deleted; pins re-baselined.
…kloads (#1278) Each workload pins one memo behavior through its selected plan: rollup across label sets, top-k reuse, chain widths and sharing, avg as Sum / Count, and which sketches an accuracy target or a declared value sign admits. The snapshots record current behavior, including the known gaps (#1260, #1270, #1281, #1282) named in each `covers`, so their fixes show up as snapshot diffs. Existing snapshots are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014qhwq4GZoWwtfFWWr9CKkr
d744bd7 to
bd3a333
Compare
|
The rebase onto main is clean, but regenerating on main changes 32 of the 33 new snapshots, and about 15
Two workloads need changes beyond their
Could you rewrite those The regenerated snapshots are on branch A separate small issue: |
Summary
Adds 33 small, single-purpose workloads to
sketch_plan_snapshots, each pinning one memo behavior through its selected plan. Closes #1278. Stacked on #1239 (refactor/1207-tools).Review guide
crates/devtools/tests/sketch_plan_snapshots/manifest.json: the 33 new entries. Eachcoverssays what the workload pins and names the issue when the plan reflects a known gap.crates/devtools/tests/sketch_plan_snapshots.rs: one header paragraph.workloads/*.jsonand<id>.jsonsnapshots are generated or copied; skim them.rollup-quantile-nested-labels,rollup-quantile-disjoint-labels,rollup-sum-chain,stddev-nested-labels,avg-nested-labels,sql-rollup-quantile-same-window,sql-rollup-quantile-different-windowstopk-5-and-10,topk-by-job-and-global,topk-5-and-10-accuracy,topk-rate-counter,topk-sum-count-over-time,topk-sum-count-over-time-accuracyquantile-over-time-no-phase,quantile-over-time-phase-0,chains-10m-5m-3m-3m,chains-10m-5m-3m-10s,chains-10m-5m-3m-10s-mvp,sum-over-time-two-intervals,quantile-phase-0-and-30s,quantile-7m-every-5msum-avg-over-time-5m,sum-5m-avg-10m-over-time,mixed-quantilesdistinct-over-time-1pct,distinct-over-time-5pct,entropy-over-time,quantile-two-targets,distinct-two-forms,entropy-and-distinct,quantile-exact,sign-unknown,sign-non-negativeAll are on
data(10k gauge series, 15 s samples), mostly every 1m at phase 0, underraw_retained(one undermvp). The two SQL workloads useworkloads/sql-table-data.tables.json.Behavior: before → after
rollup-quantile-nested-labels(quantile by (job) (0.9, data)+quantile by (job, host) (0.9, data)):asap_plannerby hand.SummaryMergefrom the(job, host)KLL.Interfaces
None.
Algorithm changes
None.
Decisions and open questions
Verification
UPDATE_SNAPSHOTS=1 cargo test --release -p asap-devtools --test corpus_snapshots -- --ignored sketch_plans_match_snapshots, then the same withoutUPDATE_SNAPSHOTS: passes (43 workloads, ~5.6 s).covers.Follow-ups
🤖 Generated with Claude Code
https://claude.ai/code/session_014qhwq4GZoWwtfFWWr9CKkr