monitoring(testing): derive gate health from raw checks

scripts/dashboards_render_atlas.py

@@ -627,6 +627,49 @@ PLATFORM_TEST_ACTIVE_SUITES_24H = (
 PLATFORM_TEST_POINT_WINDOW = "1h"
 PLATFORM_TEST_FRESH_WINDOW = "30h"
 PLATFORM_TEST_LATEST_WINDOW = "30d"
+
+
+def platform_check_status_expr(
+    suite_matcher: str,
+    *,
+    branch_matcher: str = 'branch!=""',
+    check_matcher: str = 'check!=""',
+    status_matcher: str = 'status!=""',
+    window: str = PLATFORM_TEST_FRESH_WINDOW,
+) -> str:
+    """Return recent check gauges normalized to a status label."""
+    result_matcher = status_matcher.replace("status", "result", 1)
+    base = (
+        f'__name__=~".*_quality_gate_checks_total",{suite_matcher},'
+        f'{PLATFORM_TEST_EXPORT_FILTER},{check_matcher}'
+    )
+    build_info = (
+        f'max by (suite, branch) (last_over_time(platform_quality_gate_build_info'
+        f'{{{suite_matcher},{PLATFORM_TEST_EXPORT_FILTER},{branch_matcher}}}[{window}]))'
+    )
+    with_status = (
+        f'last_over_time({{{base},{branch_matcher},{status_matcher}}}[{window}])'
+    )
+    with_result = (
+        f'label_replace(last_over_time({{{base},{branch_matcher},{result_matcher},status=""}}'
+        f'[{window}]), "status", "$1", "result", "(.*)")'
+    )
+    inferred_status = (
+        f'(last_over_time({{{base},branch="",{status_matcher}}}[{window}]) '
+        f'* on (suite) group_left(branch) ({build_info}))'
+    )
+    inferred_result = (
+        f'(label_replace(last_over_time({{{base},branch="",{result_matcher},status=""}}'
+        f'[{window}]), "status", "$1", "result", "(.*)") '
+        f'* on (suite) group_left(branch) ({build_info}))'
+    )
+    return (
+        f"sum by (suite, branch, check, status) ("
+        f"{with_status} or {with_result} or {inferred_status} or {inferred_result}"
+        f")"
+    )
+
+
 PLATFORM_TEST_SUCCESS_RATE_SUITE_TARGETS = [
     {
         "refId": chr(ord("A") + index),
@@ -679,18 +722,26 @@ PLATFORM_TEST_CHECKS_SELECTOR = (
     f"{PLATFORM_TEST_EXPORT_FILTER}"
 )
 PLATFORM_TEST_CHECK_ROLLUP_MATCHERS = (
-    f'suite=~"{PLATFORM_TEST_SUITE_CANONICAL_MATCHER}",branch!="",check!=""'
+    f'suite=~"{PLATFORM_TEST_SUITE_CANONICAL_MATCHER}"'
 )
-# This recording rule already folds recent scrape freshness into the current sample.
+PLATFORM_TEST_PRIMARY_BRANCH_MATCHER = 'branch=~"main|master|origin/main|origin/master"'
-# Do not wrap it in a multi-hour range for "latest" panels, or old failures linger.
+PLATFORM_TEST_CHECK_ROLLUP_SELECTOR = platform_check_status_expr(
-PLATFORM_TEST_CHECK_ROLLUP_SELECTOR = (
+    PLATFORM_TEST_CHECK_ROLLUP_MATCHERS,
-    f'platform_quality:check_status:present_1h{{{PLATFORM_TEST_CHECK_ROLLUP_MATCHERS},status!=""}}'
+    branch_matcher=PLATFORM_TEST_PRIMARY_BRANCH_MATCHER,
 )
 PLATFORM_TEST_CHECK_ROLLUP_OK_SELECTOR = (
-    f'platform_quality:check_status:present_1h{{{PLATFORM_TEST_CHECK_ROLLUP_MATCHERS},status=~"{PLATFORM_TEST_NON_FAILURE_STATUS}"}}'
+    platform_check_status_expr(
+        PLATFORM_TEST_CHECK_ROLLUP_MATCHERS,
+        branch_matcher=PLATFORM_TEST_PRIMARY_BRANCH_MATCHER,
+        status_matcher=f'status=~"{PLATFORM_TEST_NON_FAILURE_STATUS}"',
+    )
 )
 PLATFORM_TEST_CHECK_ROLLUP_FAILED_SELECTOR = (
-    f'platform_quality:check_status:present_1h{{{PLATFORM_TEST_CHECK_ROLLUP_MATCHERS},status!~"{PLATFORM_TEST_NON_FAILURE_STATUS}"}}'
+    platform_check_status_expr(
+        PLATFORM_TEST_CHECK_ROLLUP_MATCHERS,
+        branch_matcher=PLATFORM_TEST_PRIMARY_BRANCH_MATCHER,
+        status_matcher=f'status!~"{PLATFORM_TEST_NON_FAILURE_STATUS}"',
+    )
 )
 PLATFORM_TEST_CHECK_ROLLUP_SEEN_FLAGS = (
     f'clamp_max(max by (suite, check) (({PLATFORM_TEST_CHECK_ROLLUP_SELECTOR}) > 0), 1)'
@@ -3776,9 +3827,11 @@ def build_jobs_dashboard():
         check_matcher: str = 'check!=""',
         status_matcher: str = 'status!=""',
     ) -> str:
-        selector = (
+        selector = platform_check_status_expr(
-            f'platform_quality:check_status:present_1h{{suite=~"{suite_var}",branch!="",'
+            f'suite=~"{suite_var}"',
-            f'branch=~"{branch_var}",{check_matcher},{status_matcher}}}'
+            branch_matcher=f'branch!="",branch=~"{branch_var}"',
+            check_matcher=check_matcher,
+            status_matcher=status_matcher,
         )
         return selector
 

scripts/tests/test_dashboards_render_atlas.py

@@ -227,8 +227,11 @@ def test_jobs_dashboard_separates_current_gate_health_from_reliability():
 
     current_gate_expr = panels_by_title["Latest Gate Checks Passing by Suite"]["targets"][0]["expr"]
     assert 'check)' in current_gate_expr
-    assert "platform_quality:check_status:present_1h" in current_gate_expr
+    assert "platform_quality:check_status:present_1h" not in current_gate_expr
-    assert "platform_quality_gate_checks_total" not in current_gate_expr
+    assert '.*_quality_gate_checks_total' in current_gate_expr
+    assert "last_over_time" in current_gate_expr
+    assert 'label_replace' in current_gate_expr
+    assert 'result=~"ok|passed|success|not_applicable|skipped|na|n/a"' in current_gate_expr
     assert 'status=~"ok|passed|success|not_applicable|skipped|na|n/a"' in current_gate_expr
     assert 'status!~"ok|passed|success|not_applicable|skipped|na|n/a"' in current_gate_expr
     assert "unless on(suite, check)" in current_gate_expr
@@ -337,8 +340,10 @@ def test_jobs_dashboard_collapses_heavy_drilldowns_for_light_first_paint():
     assert failure_rate_panel["fieldConfig"]["defaults"]["max"] == 100
     assert failure_rate_panel["fieldConfig"]["defaults"]["thresholds"]["steps"][0]["color"] == "dark-blue"
     assert "increase(" not in failure_rate_panel["targets"][0]["expr"]
-    assert "platform_quality:check_status:present_1h" in failure_rate_panel["targets"][0]["expr"]
+    assert "platform_quality:check_status:present_1h" not in failure_rate_panel["targets"][0]["expr"]
-    assert "platform_quality_gate_checks_total" not in failure_rate_panel["targets"][0]["expr"]
+    assert '.*_quality_gate_checks_total' in failure_rate_panel["targets"][0]["expr"]
+    assert "last_over_time" in failure_rate_panel["targets"][0]["expr"]
+    assert 'label_replace' in failure_rate_panel["targets"][0]["expr"]
     assert "0 *" in failure_rate_panel["targets"][0]["expr"]
     assert "and on(suite)" not in failure_rate_panel["targets"][0]["expr"]