fix: warm engine resolves DDSketch _quantile suffix at query time - #110
Merged
Merged
Conversation
…PromQL queries The agent's DDSketch processor renames raw input metrics (`http_latency_ms`) to a sketched-form wire name (`http_latency_ms_quantile`) before emitting to the warm tier — by design, so warm-tier routing distinguishes raw from sketched form. But the replay client / PromQL caller still references the conceptual unsuffixed metric in `quantile_over_time(q, X[range])`, so the warm engine sees a query for `X` while its sketch store only holds `X_quantile`. Result: the lookup misses and the engine returns `status=error` to a query that is logically answerable (matchers + capability dispatch already work given the right name, per PR #108). This change picks Option 1 from the design alternatives — backend self-heals stale routing tables — because it requires no controller change and keeps the wire-side rename convention intact: * New helper `resolve_quantile_metric_alias` in `SimpleEngine` parses the query, classifies its shape, and rewrites `<metric>` → `<metric>_quantile` only when (a) shape is `Quantile` (only shapes whose data lives behind the rename), (b) the bare metric is NOT registered locally but the suffixed variant IS, and (c) the rewrite re-parses cleanly. Deployments without the rename pass through unchanged. * `handle_query_promql` calls the resolver once at entry so every downstream stage (pattern match, `QueryConfig` lookup, capability matching, `StoreQueryParams.metric`, schema label lookup) sees the same name. * `replace_metric_token` does identifier-aware substitution ([A-Za-z0-9_:] boundary check) so a hypothetical `http_latency_ms_total` elsewhere in the query is left alone. Three new regression tests in `warm_engine_replay_regression_tests` pin the contract: 1. `quantile_over_time_resolves_unsuffixed_metric_to_quantile_state` — the production failure case from ProjectASAP/ASAPCollector#46: query `quantile_over_time(0.99, http_latency_ms[1m])` against a store that only holds `http_latency_ms_quantile` returns the DDSketch quantile, not `status=error`. 2. `non_quantile_query_does_not_rewrite_metric` — `sum_over_time` against the same store does NOT rewrite (shape gating). 3. `quantile_query_without_ingest_rename_passes_through` — when the bare metric IS registered locally (deployment without rename), the resolver leaves the query untouched. Refs: ProjectASAP/ASAPCollector#46 Builds on: PR #102 (DDSketch in `compatible_agg_types(Quantile)`), PR #108 (warm-tier matcher coverage tests). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
zzylol
force-pushed
the
fix/warm-engine-resolves-quantile-suffix
branch
from
May 8, 2026 18:46
1da79e7 to
159003b
Compare
zzylol
added a commit
that referenced
this pull request
May 8, 2026
…s status=error) (#111) The MVP demo's `replay.jsonl` showed the warm-tier engine returning `status=error` for nearly every query shape — quantile_over_time (686/686), instant sum (343/343), count-of-HLL (343/343), topk (342/342), frequency (342/342). PR #110's `_quantile` alias resolver was correct in isolation but the production deploy hits a downstream mismatch this PR diagnoses and fixes. Root cause: the production warm-tier launches with `--streaming-config` only and no `--config`, so `inference_config.schema = PromQLSchema::new()` (empty). Two engine paths blow up on this: 1. `build_query_requirements_promql` falls back to `KeyByLabelNames::empty()` for `req.grouping_labels` → strict `labels_compatible` mismatches every agg config's `[zone]` → capability miss → `handle_query_promql` returns `None` → `format_unsupported_query_response` writes `status=error`. 2. `build_promql_execution_context_tail` had its own schema lookup (`schema.get_labels(metric)`) that returned `None` for empty schemas — even after capability matching succeeded — and short- circuited the entire query. Fixes: - New `SimpleEngine::resolve_metric_labels(metric)` helper that prefers the user-supplied schema and falls back to a deterministic union of `grouping_labels` across every `StreamingConfig` agg matching the metric. Both call sites above now use it, so the schema-empty deploy auto-derives sane labels from the controller- pushed agg configs. - `labels_compatible` relaxed from strict-eq to subset (`req ⊆ config`). The TODO on this function asked for exactly this; the MVP demo's `count(metric)` / `topk(K, metric)` queries with no `by (...)` modifier translate to `req.grouping_labels = []` and must match a `[zone]` agg via the merge path. Direction is asymmetric: `req ⊃ config` is still rejected (engine cannot invent partitions). The pre-existing `label_strict_superset_rejected` test is replaced with `label_superset_config_accepts_subset_query` + `label_subset_config_rejects_superset_query` to pin both halves. - `compatible_agg_types(Statistic::Count)` now lists `HLL`. The HLL accumulator answers `Statistic::Count` as a cardinality alias (`hll_sketch_accumulator.rs:220`); without HLL in the compat list every `count(<HLL-metric>)` capability-missed. - `compatible_agg_types(Statistic::Topk)` now lists `CountSketch`. Symmetric with the Count fix — `count_sketch_accumulator.rs:284` answers `Statistic::Topk`. (Note: a standalone CountSketch agg still fails downstream because `is_multi_population_value_type` requires a paired SetAggregator. The right structural fix for `top_endpoint_qps` is for the controller to plan it as `CountMinSketchWithHeap`; PR #344 declares that capability on the controller side. The engine-side test is `#[ignore]`'d with a follow-up note.) Tests: - `production_conditions_quantile_over_time_does_not_error` - `production_conditions_sum_by_zone_instant_does_not_error` - `production_conditions_count_against_hll_does_not_error` - `production_conditions_topk_against_count_sketch_does_not_error` (`#[ignore]`'d — follow-up; see comment) - `production_conditions_rate_against_cms_does_not_error` (`#[ignore]`'d — CountMinSketchAccumulator has no `Statistic::Rate` answer yet) - `label_superset_config_accepts_subset_query` / `label_subset_config_rejects_superset_query` Each new test seeds `inference_config.schema = PromQLSchema::new()` to mirror the production warm-tier deploy shape. Refs ProjectASAP/ASAPCollector#46. Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes the metric-name mismatch between the agent's DDSketch processor (renames
http_latency_ms→http_latency_ms_quantileon the wire by design) and the warm engine's lookup. PromQL callers reference the conceptual unsuffixed name; the warm engine looks it up directly and misses →status=error.Picked Option 1 (backend self-heals) over alternatives — no controller change needed, wire-side rename convention preserved.
New helper
resolve_quantile_metric_aliasinSimpleEnginerewrites<metric>→<metric>_quantileonly when:Quantile(only shape whose data lives behind the rename)Identifier-aware substitution via
[A-Za-z0-9_:]boundary sohttp_latency_ms_totalelsewhere is left alone.Test plan
cargo test -p query_engine_rust --lib— all targeted tests pass; pre-existing failures unchangedquantile_over_time_resolves_unsuffixed_metric_to_quantile_statereproduces the production failurenon_quantile_query_does_not_rewrite_metricpins shape gatingquantile_query_without_ingest_rename_passes_throughpins backwards-compat for deployments without renameRefs: ProjectASAP/ASAPCollector#46, builds on PR #102 + PR #108.
🤖 Generated with Claude Code