fix(multinode): WARMUP_S=60 + grouping_labels + queries-e2e wave queries

deploy/mvp-multinode/configs/asap/mvp-workload.yaml

@@ -51,6 +51,8 @@
   query_string: "quantile_over_time(0.99, http_requests_total_latency_ms[30s])"
   accuracy_sla: 0.01
   assign_to_role: agent
+  grouping_labels:
+  - zone
 
 # 2. LABEL AGGREGATION ACROSS SERIES AT ONE TIMESTAMP
 #    Outer: sum-by(zone) at instant. Inner: identity.
@@ -60,6 +62,8 @@
   query_string: "sum by (zone) (http_requests_total)"
   accuracy_sla: 0.0
   assign_to_role: gateway
+  grouping_labels:
+  - zone
 
 # 3. COMBINED WINDOW + LABEL AGGREGATION
 #    Inner: rate per series over 5m. Outer: sum-by(zone).
@@ -69,6 +73,8 @@
   query_string: "sum by (zone) (rate(http_requests_total[5m]))"
   accuracy_sla: 0.01
   assign_to_role: agent
+  grouping_labels:
+  - zone
 
 # 4. AD-HOC COLD-FALLBACK PROBE — exercises Gorilla archive tier
 #    Metric is configured StorageBackend::GorillaS3 in
@@ -87,6 +93,8 @@
   query_string: "count(http_requests_total{zone=\"z0\"})"
   accuracy_sla: 0.0
   assign_to_role: archive
+  grouping_labels:
+  - zone
 
 # ── Five-sketch MVP coverage (issue #46) ──────────────────────────
 #
@@ -122,6 +130,8 @@
   query_string: "quantile_over_time(0.99, request_size_bytes[30s])"
   accuracy_sla: 0.05
   assign_to_role: agent
+  grouping_labels:
+  - zone
   sketch_family_override: KLL
   target_path: warm
 
@@ -133,6 +143,8 @@
   query_string: "count(unique_users_per_min)"
   accuracy_sla: 0.02
   assign_to_role: agent
+  grouping_labels:
+  - zone
   sketch_family_override: HLL
   target_path: warm
 
@@ -145,6 +157,8 @@
   query_string: "topk(5, top_endpoint_qps)"
   accuracy_sla: 0.05
   assign_to_role: agent
+  grouping_labels:
+  - zone
   sketch_family_override: CountSketch
   target_path: warm
 
@@ -158,5 +172,7 @@
   query_string: "rate(endpoint_request_freq[5m])"
   accuracy_sla: 0.05
   assign_to_role: agent
+  grouping_labels:
+  - zone
   sketch_family_override: CountMinSketch
   target_path: warm

deploy/mvp-multinode/topology.env

@@ -58,7 +58,16 @@ EXPORTER_FIVE_SKETCH="${EXPORTER_FIVE_SKETCH:-off}"
 N_PRODUCERS_PER_NODE="${N_PRODUCERS_PER_NODE:-5}"
 
 # ── Soak / replay timing.
-WARMUP_S="${WARMUP_S:-30}"
+# WARMUP_S=60 (was 30) gives the agent's first 30s tumbling window time to
+# close and its sketch state to be flushed to the backend before replay
+# starts. With WARMUP_S=30 the first ~14 seconds of soak fired before any
+# window had closed → backend had no sketch state → asap-arm queries with
+# `quantile_over_time(...[5m])` / `sum_over_time(...[5m])` / instant
+# `sum by (zone) (http_requests_total)` returned 14 leading "error/200"s.
+# `rate(...[5m])` queries survived only because the 5-min lookback eventually
+# found a closed window. 60s = single tumble + headroom; bump to 90 if a
+# future run still shows >3 leading failures on the same queries.
+WARMUP_S="${WARMUP_S:-60}"
 SOAK_S="${SOAK_S:-90}"
 
 # ── Image set (must be loaded on every node).

deploy/mvp-singlenode/scripts/queries-e2e.json

@@ -9,7 +9,7 @@
     },
     {
         "kind": "quantile",
-        "metricsql": "quantile by (zone) (0.99, http_requests_total_latency_ms)"
+        "metricsql": "max by (zone) (quantile_over_time(0.99, http_requests_total_latency_ms[5m]))"
     },
     {
         "kind": "sum",