eval(deploy): mirror inference patterns + align queries-e2e to engine matcher

PROGRESS.md

@@ -2,6 +2,21 @@
 
 _Last updated: 2026-05-05._
 
+## Deploy-side inference mirror + queries-e2e alignment (2026-05-05)
+
+Mirrored `ASAPQuery-backend` PR #79's 33-entry warm-tier pattern set
+into the 5 deploy `backend-inference{,-cms,-cs,-hll,-kll}.yaml`
+overlays (was 1–8 entries each; now 33 / 14 / 16 / 8 / 16) and
+swapped `deploy/scripts/queries-e2e.json` off the engine-rejected
+`histogram_quantile(...)` shape onto `quantile_over_time(φ,
+*_quantile[1m])` (live-verified per "Single-pipeline multi-sketch +
+delta + queryable warm tier (2026-05-01)" above). Closes the YAML
+side of the E0 exit criterion (1) — replay queries now exact-match
+warm-tier entries instead of falling through to the cold tier. The
+engine pattern-matcher itself does not yet cover `histogram_quantile`;
+that's a separate ASAPQuery-backend PR (see
+`deploy/README.md` "Inference dispatch").
+
 ## Cross-language byte-format parity, 5/5 sketches (2026-05-05)
 
 Closes [#243](https://github.com/ProjectASAP/ASAPCollector/issues/243).

deploy/README.md

@@ -85,6 +85,59 @@ docker compose -f deploy/docker-compose/base.yml \
                -f deploy/docker-compose/baseline-b3-delta.yml up
 ```
 
+### Inference dispatch (warm-tier query coverage)
+
+The backend's "warm tier" (precomputed-sketch path) only answers
+queries whose canonical PromQL string exact-matches an entry in the
+mounted inference YAML. Five overlays live in `deploy/configs/`:
+
+| YAML | Mounted by | Covers (PromQL families × ranges) | Entries |
+|---|---|---|---|
+| `backend-inference.yaml` | `e2e-overlay.yml`, `queryengine-overlay.yml` (default) | All 33 patterns from `ASAPQuery-backend` PR #79: spatial multi-quantile, `quantile_over_time(φ ∈ {0.5, 0.9, 0.95, 0.99}, …[1m\|2m\|5m])`, `sum_over_time` / `count_over_time` × wider ranges, `rate` / `increase`, spatial `count` / `sum` / `avg`, `topk(5\|10\|50, …)`. | 33 |
+| `backend-inference-cms.yaml` | `e2e-overlay-cms.yml` | CountMinSketch families: `{sum, count, avg}`, `{sum_over_time, count_over_time, rate, increase}` × `[1m, 2m, 5m]`. | 14 |
+| `backend-inference-cs.yaml` | `e2e-overlay-cs.yml` | CountSketch families: same as CMS plus `topk(5\|10\|50, …)`. | 16 |
+| `backend-inference-hll.yaml` | `e2e-overlay-hll.yml` | HLL cardinality families: spatial `count(metric_hll)` and `count_over_time(metric_hll[1m\|2m\|5m])` for both the counter (`http_requests_total_hll`) and gauge (`http_requests_total_latency_ms_hll`) flavours. | 8 |
+| `backend-inference-kll.yaml` | `e2e-overlay-kll.yml` | KLL rank-quantile families: spatial `quantile by (zone) (φ, …)` and `quantile_over_time(φ, metric_kll[1m\|2m\|5m])` × `φ ∈ {0.5, 0.9, 0.95, 0.99}`. | 16 |
+
+This is the canonical paper-experiment pattern set (PR #79
+`tests/inference_yaml_pattern_coverage.rs` is the runtime contract).
+Adding a new query family here without a matching entry in PR #79's
+canonical YAML risks shipping warm-tier "promises" the engine can't
+keep — the YAML is checked exact-string at request time by
+`find_query_config`.
+
+#### Metric-name conventions per overlay
+
+| Overlay | Backend-side metric name | Why |
+|---|---|---|
+| `backend-inference.yaml` | `http_requests_total_latency_ms_quantile` (KLL/DDSketch quantile patterns), `http_requests_total` (CMS/CountSketch/HLL Sum/Count/Topk patterns) | Default deploy uses `gateway-aggregate-from-raw.yaml` → DDSketch → `metric_suffix: "_quantile"`. Sum/Count/Topk patterns target the raw counter forwarded unsuffixed by CMS/CountSketch processors (see `sketchcol-agent-{cms,cs}-direct.yaml`). |
+| `backend-inference-cms.yaml`, `-cs.yaml` | `http_requests_total` | CMS / CountSketch direct agents preserve the raw metric name (no `metric_suffix`). |
+| `backend-inference-hll.yaml` | `http_requests_total_hll`, `http_requests_total_latency_ms_hll` | `sketchcol-agent-hll-direct.yaml` adds `metric_suffix: "_hll"`. |
+| `backend-inference-kll.yaml` | `http_requests_total_latency_ms_kll` | `sketchcol-agent-kll-direct.yaml` adds `metric_suffix: "_kll"`. |
+
+#### Open gap: `histogram_quantile(φ, …)` is NOT covered
+
+The engine pattern matcher
+(`asap-query-engine/src/engines/simple_engine.rs::controller_patterns`)
+includes `quantile_over_time` and the spatial `quantile by (…)` ops
+but does **not** include a `histogram_quantile` pattern block.
+`histogram_quantile(0.99, sum by (le) (http_requests_total_latency_ms))`
+fails the engine matcher before reaching `find_query_config`, even
+when an entry of that exact string is present in the YAML.
+
+Two paths forward (in priority order):
+
+1. **Today (this PR's choice for `queries-e2e.json`):** use the
+   pre-aggregated `quantile_over_time(φ, *_quantile[…])` shape. The
+   gateway / agent path produces `http_requests_total_latency_ms_quantile`
+   (DDSketch / KLL backed) which the engine matches and answers.
+   PROGRESS.md "Single-pipeline multi-sketch + delta + queryable
+   warm tier (2026-05-01)" verified this path live (`q=0.5 →
+   19.49`).
+2. **Tomorrow (separate engine PR):** extend `controller_patterns`
+   with a `histogram_quantile` block. Out of scope for E0; tracked
+   alongside PR #79 follow-ups.
+
 ### E0: single-cell smoke (P5–P9 end-to-end against a live stack)
 
 Smallest cell that exercises the whole P1–P9 path. Use this
@@ -126,6 +179,11 @@ python3 deploy/scripts/plan_transition.py \
     --transition-out /tmp/cell-smoke-e0/transition.jsonl \
     --sample-out /tmp/cell-smoke-e0/sample.jsonl \
     --soak-secs 60 --pre-transition-secs 20 &
+# Note: the `histogram_quantile(...)` shape above is INTENTIONALLY
+# unmatched by the engine — it's the capability-miss probe used by
+# `plan_transition.py` to drive a fresh plan publish. Replay-side
+# queries (`queries-e2e.json`) deliberately use shapes that DO match
+# (see "Inference dispatch" above).
 wait
 
 # 4. Snapshot ground truth from the cold-store volume (the volume

deploy/configs/backend-inference-cms.yaml

@@ -1,3 +1,27 @@
+# Warm-tier dispatch table (CountMinSketch single-sketch overlay).
+#
+# Subset of `backend-inference.yaml` (which mirrors PR #79's
+# `inference_config.yaml`) restricted to the families that route to
+# the CountMinSketch accumulator:
+#
+#   * Spatial `sum` / `count` — Statistic::{Sum,Count} no-key
+#     fallback returns the min-row sum (canonical CMS total-event
+#     estimator). PROGRESS.md "All-five-sketch runtime e2e
+#     verification (2026-04-30)" verified
+#     `sum_over_time(http_requests_total[1m])` → 145735 with ε≈0.0027.
+#   * `sum_over_time` / `count_over_time` over [1m]/[2m]/[5m] —
+#     Statistic::Sum / Statistic::Count via
+#     `additive_frequency` accuracy envelope.
+#   * `rate` / `increase` — same family (delta arithmetic over Sum).
+#
+# `topk` / `histogram_quantile` / `quantile_over_time` excluded
+# (CMS does not back those statistics — see the
+# `compatible_agg_types` table in `capability_matching.rs`).
+#
+# Metric: `http_requests_total` (no suffix). The CMS direct agent
+# overlay (`sketchcol-agent-cms-direct.yaml`) uses
+# `metric_name: "http_requests_total"` and preserves the raw name
+# end-to-end.
 cleanup_policy:
   name: "circular_buffer"
 metrics:
@@ -7,11 +31,62 @@ metrics:
   - node
   - pod
 queries:
+# ── Spatial Sum / Count / Avg ─────────────────────────────────────────
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: sum(http_requests_total)
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count(http_requests_total)
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: avg(http_requests_total)
+# ── sum_over_time / count_over_time × wider ranges ────────────────────
 - aggregations:
   - aggregation_id: 1
     num_aggregates_to_retain: 6
   query: sum_over_time(http_requests_total[1m])
 - aggregations:
   - aggregation_id: 1
     num_aggregates_to_retain: 6
-  query: sum(http_requests_total)
+  query: sum_over_time(http_requests_total[2m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: sum_over_time(http_requests_total[5m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total[1m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total[2m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total[5m])
+# ── rate / increase (delta arithmetic over Sum family) ────────────────
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: rate(http_requests_total[1m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: rate(http_requests_total[2m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: rate(http_requests_total[5m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: increase(http_requests_total[1m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: increase(http_requests_total[5m])

deploy/configs/backend-inference-cs.yaml

@@ -1,3 +1,26 @@
+# Warm-tier dispatch table (CountSketch single-sketch overlay).
+#
+# Subset of `backend-inference.yaml` (which mirrors PR #79's
+# `inference_config.yaml`) restricted to the families that route to
+# the CountSketch accumulator:
+#
+#   * `topk(N, …)` — Statistic::Topk; CountSketch's `query_statistic`
+#     returns the row-mean total when no key is supplied (per PR #79
+#     wiring; per-key enumeration needs a paired SetAggregator on the
+#     agent — tracked upstream).
+#   * Spatial `sum` / `count` — Statistic::{Sum,Count} no-key
+#     fallback (PROGRESS.md "All-five-sketch runtime e2e verification
+#     (2026-04-30)" — `sum_over_time(http_requests_total[1m])` →
+#     24266 with ε=0.03 `additive_frequency` envelope).
+#   * `sum_over_time` / `count_over_time` × [1m]/[2m]/[5m].
+#   * `rate` / `increase` — same Sum family.
+#
+# `quantile_over_time` / `histogram_quantile` excluded (CountSketch
+# does not back rank statistics).
+#
+# Metric: `http_requests_total` (no suffix). The CountSketch direct
+# agent overlay (`sketchcol-agent-cs-direct.yaml`) preserves the raw
+# metric name through the processor.
 cleanup_policy:
   name: "circular_buffer"
 metrics:
@@ -6,17 +29,72 @@ metrics:
   - rack
   - node
   - pod
-  http_requests_total_latency_ms:
-  - zone
-  - rack
-  - node
-  - pod
 queries:
+# ── Spatial Sum / Count ───────────────────────────────────────────────
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: sum(http_requests_total)
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count(http_requests_total)
+# ── sum_over_time / count_over_time × wider ranges ────────────────────
 - aggregations:
   - aggregation_id: 1
     num_aggregates_to_retain: 6
   query: sum_over_time(http_requests_total[1m])
 - aggregations:
-  - aggregation_id: 2
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: sum_over_time(http_requests_total[2m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: sum_over_time(http_requests_total[5m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total[1m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total[2m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total[5m])
+# ── rate / increase (delta arithmetic over Sum family) ────────────────
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: rate(http_requests_total[1m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: rate(http_requests_total[2m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: rate(http_requests_total[5m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: increase(http_requests_total[1m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: increase(http_requests_total[5m])
+# ── Top-K (CountSketch heavy-hitter readout) ──────────────────────────
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: topk(5, http_requests_total)
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: topk(10, http_requests_total)
+- aggregations:
+  - aggregation_id: 1
     num_aggregates_to_retain: 6
-  query: sum_over_time(http_requests_total_latency_ms[1m])
+  query: topk(50, http_requests_total)

deploy/configs/backend-inference-hll.yaml

@@ -1,3 +1,24 @@
+# Warm-tier dispatch table (HyperLogLog single-sketch overlay).
+#
+# Subset of `backend-inference.yaml` (which mirrors PR #79's
+# `inference_config.yaml`) restricted to the families that route to
+# the HLL accumulator:
+#
+#   * `count(metric)` — Statistic::Count; HLL's `query_statistic`
+#     accepts both `Cardinality` and `Count` (alias) and returns the
+#     unique-cardinality estimate. PROGRESS.md "All-five-sketch
+#     runtime e2e verification (2026-04-30)" verified
+#     `count(http_requests_total)` → 149.68 distinct, ε=0.008
+#     `relative_cardinality` envelope.
+#   * `count_over_time(metric[…])` — temporal Count.
+#
+# `sum` / `quantile_over_time` / `topk` / `rate` excluded (HLL is a
+# cardinality sketch — no value semantics).
+#
+# Metric: `http_requests_total_hll` / `http_requests_total_latency_ms_hll`
+# — the HLL direct agent overlay (`sketchcol-agent-hll-direct.yaml`)
+# uses `metric_suffix: "_hll"`, so the backend-side metric name
+# is the suffixed form.
 cleanup_policy:
   name: "circular_buffer"
 metrics:
@@ -12,15 +33,37 @@ metrics:
   - node
   - pod
 queries:
+# ── Spatial Cardinality (count over an HLL-backed metric) ─────────────
 - aggregations:
   - aggregation_id: 1
     num_aggregates_to_retain: 6
   query: count(http_requests_total_hll)
+- aggregations:
+  - aggregation_id: 2
+    num_aggregates_to_retain: 6
+  query: count(http_requests_total_latency_ms_hll)
+# ── count_over_time × wider ranges ────────────────────────────────────
 - aggregations:
   - aggregation_id: 1
     num_aggregates_to_retain: 6
   query: count_over_time(http_requests_total_hll[1m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total_hll[2m])
+- aggregations:
+  - aggregation_id: 1
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total_hll[5m])
 - aggregations:
   - aggregation_id: 2
     num_aggregates_to_retain: 6
   query: count_over_time(http_requests_total_latency_ms_hll[1m])
+- aggregations:
+  - aggregation_id: 2
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total_latency_ms_hll[2m])
+- aggregations:
+  - aggregation_id: 2
+    num_aggregates_to_retain: 6
+  query: count_over_time(http_requests_total_latency_ms_hll[5m])