Increase frozen perf compute budget by 50%

.github/workflows/scenario-dev.yml

@@ -81,8 +81,8 @@ jobs:
       DUCKGRES_SCENARIO_MAX_RUNTIME: 4h
       DUCKGRES_SCENARIO_GO_TEST_TIMEOUT: 4h15m
       # Add process headroom for repeated full-dataset pgwire aggregates.
-      DUCKGRES_K8S_WORKER_CPU_REQUEST: "2"
-      DUCKGRES_K8S_WORKER_MEMORY_REQUEST: 8Gi
+      DUCKGRES_K8S_WORKER_CPU_REQUEST: "3"
+      DUCKGRES_K8S_WORKER_MEMORY_REQUEST: 12Gi
     steps:
       - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
 

tests/mw-dev/README.md

@@ -54,7 +54,7 @@ The isolated control plane's default worker request is configurable through
 `DUCKGRES_K8S_WORKER_CPU_REQUEST` and
 `DUCKGRES_K8S_WORKER_MEMORY_REQUEST`; `run.sh` defaults them to `750m` and
 `1536Mi`, respectively, preserving the e2e harness's worker-packing behavior.
-`scenario-dev.yml` explicitly overrides them to 2 CPU and 8Gi for the frozen
+`scenario-dev.yml` explicitly overrides them to 3 CPU and 12Gi for the frozen
 perf workload. Direct `run.sh` callers can make the same explicit override.
 
 ### Scenario Trino readiness
@@ -121,7 +121,7 @@ Pod Identity associations.
 ## Isolated Trino lane
 
 The Trino lane never uses the shared mw-dev Trino namespace or coordinator.
-It deploys one coordinator, two workers, and an OPA sidecar in the lane's
+It deploys one coordinator, three workers, and an OPA sidecar in the lane's
 `duckgres-ci-pr-<N>` namespace. The control plane projects fixed-name auth,
 tenant-secret, OPA-token, and resource-group objects into that same namespace;
 its cell id and PostgreSQL catalog store are also PR-local. This isolation is
@@ -132,12 +132,13 @@ The lane defaults `TRINO_IMAGE` to the pinned PostHog fork used when this suite
 was added. That fork contains the DuckLake connector and PostgreSQL dynamic
 catalog store; upstream `trinodb/trino` is not compatible. Update the default
 in `run.sh` and `e2e-mw-dev.yml` together when promoting a Trino build.
-Each Trino worker has requests and limits of 1 CPU and 4Gi. Together they match
-the frozen perf Duckgres worker's aggregate 2 CPU and 8Gi execution budget while
-exercising Trino's distributed execution path. Trino permits 2GB of query
-memory per worker and 4GB cluster-wide. The coordinator does not execute query
-tasks (`node-scheduler.include-coordinator=false`) and is additional Trino
-control-plane overhead rather than part of the matched execution budget.
+Each Trino worker has requests and limits of 1 CPU and 4Gi. Together they
+match the frozen perf Duckgres worker's aggregate 3 CPU and 12Gi execution
+budget while exercising Trino's distributed execution path. Trino permits 2GB
+of query memory per worker and 6GB cluster-wide. The coordinator does not
+execute query tasks (`node-scheduler.include-coordinator=false`) and is
+additional Trino control-plane overhead rather than part of the matched
+execution budget.
 `TRINO_TLS_PASSWORD` defaults to `duckgres-e2e-keystore`; it protects only the
 random, two-day, per-run PKCS12 file. `run.sh` generates a fresh CA and leaf
 certificate under `DUCKGRES_CI_SECRET_DIR`, mounts the CA into the PR control

tests/mw-dev/manifests.trino.tmpl.yaml

@@ -52,7 +52,7 @@ data:
     http-server.https.keystore.path=/etc/trino/tls/keystore.p12
     http-server.https.keystore.key=${ENV:TRINO_TLS_KEYSTORE_PASSWORD}
     discovery.uri=http://duckgres-trino.${NAMESPACE}.svc:8080
-    query.max-memory=4GB
+    query.max-memory=6GB
     # The coordinator does not execute tasks and retains headroom in its 2G
     # heap; execution workers carry the 2GB per-node allowance below.
     query.max-memory-per-node=1GB
@@ -102,7 +102,7 @@ data:
     coordinator=false
     http-server.http.port=8080
     discovery.uri=http://duckgres-trino.${NAMESPACE}.svc:8080
-    query.max-memory=4GB
+    query.max-memory=6GB
     query.max-memory-per-node=2GB
     shutdown.grace-period=10s
     catalog.management=dynamic
@@ -262,7 +262,7 @@ spec:
             periodSeconds: 5
           resources:
             requests: { cpu: "1", memory: 4Gi }
-            # Two replicas provide the same aggregate 2-CPU/8Gi execution
+            # Three replicas provide the same aggregate 3-CPU/12Gi execution
             # budget as the frozen-perf Duckgres worker. Keep requests equal
             # to limits so node contention cannot change the comparison.
             limits: { cpu: "1", memory: 4Gi }

tests/mw-dev/run.sh

@@ -346,12 +346,13 @@ cmd_deploy() {
     # Patch the Deployment resources directly. The CI deployer intentionally
     # cannot patch the deployments/scale subresource, while it already needs
     # narrowly scoped Deployment patch access for the control-plane config.
-    # Two 1-CPU/4Gi workers match the frozen-perf Duckgres worker's aggregate
-    # 2-CPU/8Gi execution budget while exercising Trino's distributed path.
+    # Three 1-CPU/4Gi workers match the frozen-perf Duckgres worker's
+    # aggregate 3-CPU/12Gi execution budget while exercising Trino's
+    # distributed path.
     "${KUBECTL[@]}" -n "$NS" patch deployment duckgres-trino-coordinator \
       --type=merge -p '{"spec":{"replicas":1}}'
     "${KUBECTL[@]}" -n "$NS" patch deployment duckgres-trino-worker \
-      --type=merge -p '{"spec":{"replicas":2}}'
+      --type=merge -p '{"spec":{"replicas":3}}'
     "${KUBECTL[@]}" -n "$NS" rollout status deploy/duckgres-trino-coordinator --timeout=300s
     "${KUBECTL[@]}" -n "$NS" rollout status deploy/duckgres-trino-worker --timeout=300s
   fi

tests/mw-dev/run_sh_test.go

@@ -126,7 +126,7 @@ func TestTrinoDeployStartsWorkloadsWithoutScaleSubresource(t *testing.T) {
 	}
 	for deployment, replicas := range map[string]int{
 		"duckgres-trino-coordinator": 1,
-		"duckgres-trino-worker":      2,
+		"duckgres-trino-worker":      3,
 	} {
 		want := "patch deployment " + deployment + " --type=merge -p {\"spec\":{\"replicas\":" + strconv.Itoa(replicas) + "}}"
 		if !strings.Contains(calls, want) {
@@ -185,6 +185,10 @@ func TestTrinoWorkersMatchDuckgresAggregateCompute(t *testing.T) {
 			t.Errorf("Trino worker %s = %v, want 1 CPU and 4Gi", field, values)
 		}
 	}
+	workerData := workerConfig["data"].(map[string]any)
+	if jvmConfig := workerData["jvm.config"].(string); !strings.Contains(jvmConfig, "-Xmx3G") {
+		t.Errorf("Trino worker JVM config does not fit its 4Gi pod:\n%s", jvmConfig)
+	}
 
 	for name, expectedPerNode := range map[string]string{
 		"coordinator": "1GB",
@@ -196,7 +200,7 @@ func TestTrinoWorkersMatchDuckgresAggregateCompute(t *testing.T) {
 		}
 		data := manifest["data"].(map[string]any)
 		config := data["config.properties"].(string)
-		for _, want := range []string{"query.max-memory=4GB", "query.max-memory-per-node=" + expectedPerNode} {
+		for _, want := range []string{"query.max-memory=6GB", "query.max-memory-per-node=" + expectedPerNode} {
 			if !strings.Contains(config, want) {
 				t.Errorf("Trino %s config missing %q:\n%s", name, want, config)
 			}

tests/mw-dev/scenario/script_test.go

@@ -90,8 +90,8 @@ func TestDevScenarioWorkflowUsesUnifiedMwDevHarness(t *testing.T) {
 		"TRINO_POD_IDENTITY_ROLE: ${{ secrets.MW_DEV_TRINO_POD_IDENTITY_ROLE }}",
 		"TRINO_IMAGE: ghcr.io/posthog/trino:",
 		"E2E_SUITE: ${{ (github.event_name == 'schedule' || inputs.scenario == 'posthog_frozen_perf') && 'trino' || 'neutral' }}",
-		"DUCKGRES_K8S_WORKER_CPU_REQUEST: \"2\"",
-		"DUCKGRES_K8S_WORKER_MEMORY_REQUEST: 8Gi",
+		"DUCKGRES_K8S_WORKER_CPU_REQUEST: \"3\"",
+		"DUCKGRES_K8S_WORKER_MEMORY_REQUEST: 12Gi",
 		"role-duration-seconds: 16200",
 		"tests/mw-dev/run.sh deploy",
 		"tests/mw-dev/run.sh test-scenario",