monitoring: replace failure table with 24h suite pass snapshot

scripts/dashboards_render_atlas.py

@@ -515,6 +515,31 @@ PLATFORM_TEST_SUCCESS_RATE_SUITE_TARGETS = [
     {"refId": "B", "expr": PLATFORM_TEST_SUCCESS_RATE_METIS_SERIES, "legendFormat": "metis"},
     {"refId": "C", "expr": PLATFORM_TEST_SUCCESS_RATE_ANANKE_SERIES, "legendFormat": "ananke"},
 ] + PLATFORM_TEST_GENERIC_SUITE_TARGETS
+
+PLATFORM_TEST_SUCCESS_RATE_24H_NATIVE_BY_SUITE = (
+    'label_replace('
+    '(100 * (sum(increase(ariadne_task_runs_total{status="ok"}[24h]))) / clamp_min((sum(increase(ariadne_task_runs_total[24h]))), 1)) '
+    'and on() ((sum(increase(ariadne_task_runs_total[24h]))) > 0), '
+    '"suite", "ariadne", "__name__", ".*") '
+    'or label_replace('
+    '(100 * ((sum(increase(metis_builds_total{status="ok"}[24h])) + sum(increase(metis_flashes_total{status="ok"}[24h])))) '
+    '/ clamp_min(((sum(increase(metis_builds_total[24h])) + sum(increase(metis_flashes_total[24h])))), 1)) '
+    'and on() (((sum(increase(metis_builds_total[24h])) + sum(increase(metis_flashes_total[24h])))) > 0), '
+    '"suite", "metis", "__name__", ".*") '
+    'or label_replace('
+    '(100 * (sum(increase(ananke_quality_gate_runs_total{suite="ananke",status="ok"}[24h]))) '
+    '/ clamp_min((sum(increase(ananke_quality_gate_runs_total{suite="ananke"}[24h]))), 1)) '
+    'and on() ((sum(increase(ananke_quality_gate_runs_total{suite="ananke"}[24h]))) > 0), '
+    '"suite", "ananke", "__name__", ".*")'
+)
+PLATFORM_TEST_SUCCESS_RATE_24H_GENERIC_BY_SUITE = (
+    '(100 * (sum by (suite) (increase(platform_quality_gate_runs_total{status=~"ok|passed|success"}[24h]))) '
+    '/ clamp_min((sum by (suite) (increase(platform_quality_gate_runs_total[24h]))), 1)) '
+    'and on(suite) ((sum by (suite) (increase(platform_quality_gate_runs_total[24h]))) > 0)'
+)
+PLATFORM_TEST_SUCCESS_RATE_24H_BY_SUITE = (
+    f'sort_desc(({PLATFORM_TEST_SUCCESS_RATE_24H_NATIVE_BY_SUITE}) or ({PLATFORM_TEST_SUCCESS_RATE_24H_GENERIC_BY_SUITE}))'
+)
 ANANKE_SELECTOR = 'job="ananke-power"'
 ANANKE_UPS_DB_NAME = "Pyrphoros"
 ANANKE_UPS_DB_NODE = "titan-db"
@@ -1583,24 +1608,29 @@ def build_overview():
     )
     panels.append(test_success)
     panels.append(
-        table_panel(
+        bargauge_panel(
             47,
-            "Platform Test Failures (24h)",
+            "Platform Suite Pass Rate (24h)",
-            PLATFORM_TEST_FAILURES_24H_BY_SUITE,
+            PLATFORM_TEST_SUCCESS_RATE_24H_BY_SUITE,
             {"h": 5, "w": 6, "x": 18, "y": 7},
-            unit="none",
+            unit="percent",
             instant=True,
-            transformations=[
+            legend="{{suite}}",
-                {"id": "labelsToFields", "options": {}},
+            sort_order="desc",
-                {"id": "organize", "options": {"excludeByName": {"Time": True}}},
+            thresholds={
-                {"id": "sortBy", "options": {"fields": ["Value"], "order": "desc"}},
+                "mode": "absolute",
-            ],
+                "steps": [
-            options={"showHeader": True, "cellHeight": "sm"},
+                    {"color": "red", "value": None},
-            footer={"show": False},
+                    {"color": "yellow", "value": 80},
+                    {"color": "green", "value": 95},
+                ],
+            },
         )
     )
     panels[-1]["links"] = link_to("atlas-jobs")
-    panels[-1]["description"] = "Failures by suite in the last 24 hours. This is a per-suite breakdown, not a single opaque total."
+    panels[-1]["description"] = (
+        "24-hour per-suite pass-rate snapshot. This complements the 7-day trend by showing each suite's current quality posture."
+    )
 
     panels.append(
         stat_panel(

services/monitoring/dashboards/atlas-overview.json

@@ -1951,8 +1951,8 @@
     },
     {
       "id": 47,
-      "type": "table",
+      "type": "bargauge",
-      "title": "Platform Test Failures (24h)",
+      "title": "Platform Suite Pass Rate (24h)",
       "datasource": {
         "type": "prometheus",
         "uid": "atlas-vm"
@@ -1965,41 +1965,49 @@
       },
       "targets": [
         {
-          "expr": "sort_desc(sum by (suite) (label_replace(increase(ariadne_task_runs_total{status!=\"ok\"}[24h]), \"suite\", \"ariadne\", \"__name__\", \".*\") or label_replace(increase(metis_builds_total{status=\"error\"}[24h]), \"suite\", \"metis\", \"__name__\", \".*\") or label_replace(increase(metis_flashes_total{status=\"error\"}[24h]), \"suite\", \"metis\", \"__name__\", \".*\") or label_replace(increase(ananke_quality_gate_runs_total{suite=\"ananke\",status=\"failed\"}[24h]), \"suite\", \"ananke\", \"__name__\", \".*\") or increase(platform_quality_gate_runs_total{status!~\"ok|passed|success\"}[24h])))",
+          "expr": "sort_desc((label_replace((100 * (sum(increase(ariadne_task_runs_total{status=\"ok\"}[24h]))) / clamp_min((sum(increase(ariadne_task_runs_total[24h]))), 1)) and on() ((sum(increase(ariadne_task_runs_total[24h]))) > 0), \"suite\", \"ariadne\", \"__name__\", \".*\") or label_replace((100 * ((sum(increase(metis_builds_total{status=\"ok\"}[24h])) + sum(increase(metis_flashes_total{status=\"ok\"}[24h])))) / clamp_min(((sum(increase(metis_builds_total[24h])) + sum(increase(metis_flashes_total[24h])))), 1)) and on() (((sum(increase(metis_builds_total[24h])) + sum(increase(metis_flashes_total[24h])))) > 0), \"suite\", \"metis\", \"__name__\", \".*\") or label_replace((100 * (sum(increase(ananke_quality_gate_runs_total{suite=\"ananke\",status=\"ok\"}[24h]))) / clamp_min((sum(increase(ananke_quality_gate_runs_total{suite=\"ananke\"}[24h]))), 1)) and on() ((sum(increase(ananke_quality_gate_runs_total{suite=\"ananke\"}[24h]))) > 0), \"suite\", \"ananke\", \"__name__\", \".*\")) or ((100 * (sum by (suite) (increase(platform_quality_gate_runs_total{status=~\"ok|passed|success\"}[24h]))) / clamp_min((sum by (suite) (increase(platform_quality_gate_runs_total[24h]))), 1)) and on(suite) ((sum by (suite) (increase(platform_quality_gate_runs_total[24h]))) > 0)))",
           "refId": "A",
+          "legendFormat": "{{suite}}",
           "instant": true
         }
       ],
       "fieldConfig": {
         "defaults": {
-          "unit": "none",
+          "unit": "percent",
-          "custom": {
+          "min": 0,
-            "filterable": true
+          "max": 100,
+          "thresholds": {
+            "mode": "absolute",
+            "steps": [
+              {
+                "color": "red",
+                "value": null
+              },
+              {
+                "color": "yellow",
+                "value": 80
+              },
+              {
+                "color": "green",
+                "value": 95
+              }
+            ]
           }
         },
         "overrides": []
       },
       "options": {
-        "showHeader": true,
+        "displayMode": "gradient",
-        "columnFilters": false,
+        "orientation": "horizontal",
-        "cellHeight": "sm",
+        "reduceOptions": {
-        "footer": {
+          "calcs": [
-          "show": false
+            "lastNotNull"
+          ],
+          "fields": "",
+          "values": false
         }
       },
       "transformations": [
-        {
-          "id": "labelsToFields",
-          "options": {}
-        },
-        {
-          "id": "organize",
-          "options": {
-            "excludeByName": {
-              "Time": true
-            }
-          }
-        },
         {
           "id": "sortBy",
           "options": {
@@ -2017,7 +2025,7 @@
           "targetBlank": true
         }
       ],
-      "description": "Failures by suite in the last 24 hours. This is a per-suite breakdown, not a single opaque total."
+      "description": "24-hour per-suite pass-rate snapshot. This complements the 7-day trend by showing each suite's current quality posture."
     },
     {
       "id": 30,

services/monitoring/grafana-dashboard-overview.yaml

@@ -1960,8 +1960,8 @@ data:
         },
         {
           "id": 47,
-          "type": "table",
+          "type": "bargauge",
-          "title": "Platform Test Failures (24h)",
+          "title": "Platform Suite Pass Rate (24h)",
           "datasource": {
             "type": "prometheus",
             "uid": "atlas-vm"
@@ -1974,41 +1974,49 @@ data:
           },
           "targets": [
             {
-              "expr": "sort_desc(sum by (suite) (label_replace(increase(ariadne_task_runs_total{status!=\"ok\"}[24h]), \"suite\", \"ariadne\", \"__name__\", \".*\") or label_replace(increase(metis_builds_total{status=\"error\"}[24h]), \"suite\", \"metis\", \"__name__\", \".*\") or label_replace(increase(metis_flashes_total{status=\"error\"}[24h]), \"suite\", \"metis\", \"__name__\", \".*\") or label_replace(increase(ananke_quality_gate_runs_total{suite=\"ananke\",status=\"failed\"}[24h]), \"suite\", \"ananke\", \"__name__\", \".*\") or increase(platform_quality_gate_runs_total{status!~\"ok|passed|success\"}[24h])))",
+              "expr": "sort_desc((label_replace((100 * (sum(increase(ariadne_task_runs_total{status=\"ok\"}[24h]))) / clamp_min((sum(increase(ariadne_task_runs_total[24h]))), 1)) and on() ((sum(increase(ariadne_task_runs_total[24h]))) > 0), \"suite\", \"ariadne\", \"__name__\", \".*\") or label_replace((100 * ((sum(increase(metis_builds_total{status=\"ok\"}[24h])) + sum(increase(metis_flashes_total{status=\"ok\"}[24h])))) / clamp_min(((sum(increase(metis_builds_total[24h])) + sum(increase(metis_flashes_total[24h])))), 1)) and on() (((sum(increase(metis_builds_total[24h])) + sum(increase(metis_flashes_total[24h])))) > 0), \"suite\", \"metis\", \"__name__\", \".*\") or label_replace((100 * (sum(increase(ananke_quality_gate_runs_total{suite=\"ananke\",status=\"ok\"}[24h]))) / clamp_min((sum(increase(ananke_quality_gate_runs_total{suite=\"ananke\"}[24h]))), 1)) and on() ((sum(increase(ananke_quality_gate_runs_total{suite=\"ananke\"}[24h]))) > 0), \"suite\", \"ananke\", \"__name__\", \".*\")) or ((100 * (sum by (suite) (increase(platform_quality_gate_runs_total{status=~\"ok|passed|success\"}[24h]))) / clamp_min((sum by (suite) (increase(platform_quality_gate_runs_total[24h]))), 1)) and on(suite) ((sum by (suite) (increase(platform_quality_gate_runs_total[24h]))) > 0)))",
               "refId": "A",
+              "legendFormat": "{{suite}}",
               "instant": true
             }
           ],
           "fieldConfig": {
             "defaults": {
-              "unit": "none",
+              "unit": "percent",
-              "custom": {
+              "min": 0,
-                "filterable": true
+              "max": 100,
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "red",
+                    "value": null
+                  },
+                  {
+                    "color": "yellow",
+                    "value": 80
+                  },
+                  {
+                    "color": "green",
+                    "value": 95
+                  }
+                ]
               }
             },
             "overrides": []
           },
           "options": {
-            "showHeader": true,
+            "displayMode": "gradient",
-            "columnFilters": false,
+            "orientation": "horizontal",
-            "cellHeight": "sm",
+            "reduceOptions": {
-            "footer": {
+              "calcs": [
-              "show": false
+                "lastNotNull"
+              ],
+              "fields": "",
+              "values": false
             }
           },
           "transformations": [
-            {
-              "id": "labelsToFields",
-              "options": {}
-            },
-            {
-              "id": "organize",
-              "options": {
-                "excludeByName": {
-                  "Time": true
-                }
-              }
-            },
             {
               "id": "sortBy",
               "options": {
@@ -2026,7 +2034,7 @@ data:
               "targetBlank": true
             }
           ],
-          "description": "Failures by suite in the last 24 hours. This is a per-suite breakdown, not a single opaque total."
+          "description": "24-hour per-suite pass-rate snapshot. This complements the 7-day trend by showing each suite's current quality posture."
         },
         {
           "id": 30,