Fold coverage_ advisor verdicts into flaky_trunk

- Strip a leading "coverage_" from signal_key into base_key before
  normalizing, so coverage/backfill dispatches classify the same red
- Group the advisor verdict subquery by (suspect_commit, base_key)
  instead of raw signal_key
- Prefer a native (non-coverage_) verdict over a coverage_ one for the
  same (commit, job), latest timestamp breaking ties
- Apply the identical change across all four flaky_trunk queries:
  entity_runs, jobs, runner_labels, timeseries

Coverage/backfill advisor dispatches emit signal_key = 'coverage_' +
the native job signal_key. Without stripping the prefix these verdicts
normalized to a distinct adv_norm and failed to join the corresponding
trunk red, leaving those reds unclassified. No native job signal_key
starts with 'coverage_', so the strip is safe.

Signed-off-by: Jean Schmidt <contato@jschmidt.me>

Author: jeanschmidt

torchci/clickhouse_queries/flaky_trunk_entity_runs/query.sql

@@ -50,20 +50,27 @@ advisor_agg AS (
     FROM (
         SELECT
             toString(suspect_commit) AS head_sha,
+            -- Strip a leading "coverage_": coverage/backfill advisor dispatches use
+            -- signal_key = 'coverage_' + the native job signal_key, so stripping it
+            -- lets such a verdict normalize to (and classify) the same red. No native
+            -- job signal_key starts with 'coverage_'.
+            replaceRegexpOne(signal_key, '^coverage_', '') AS base_key,
             replaceRegexpOne(
                 replaceRegexpOne(
-                    signal_key,
+                    base_key,
                     ' \\[[^\\]]+\\]$', ''
                 ),
                 ', [0-9]+, [0-9]+, .+\\)', ')'
             ) AS adv_norm,
-            argMax(verdict, timestamp) AS verdict
+            -- Prefer a native (non-coverage_) verdict over a coverage_ one for the same
+            -- (commit, job); among equals, the latest timestamp wins.
+            argMax(verdict, (toUInt8(NOT startsWith(signal_key, 'coverage_')), timestamp)) AS verdict
         FROM misc.autorevert_advisor_verdicts
         WHERE
             repo = {repo: String}
             AND signal_source = 'job'
             AND timestamp >= {startTime: DateTime64(3)}
-        GROUP BY suspect_commit, signal_key
+        GROUP BY suspect_commit, base_key
     )
     GROUP BY head_sha, adv_norm
 ),

torchci/clickhouse_queries/flaky_trunk_jobs/query.sql

@@ -55,20 +55,27 @@ advisor_agg AS (
     FROM (
         SELECT
             toString(suspect_commit) AS head_sha,
+            -- Strip a leading "coverage_": coverage/backfill advisor dispatches use
+            -- signal_key = 'coverage_' + the native job signal_key, so stripping it
+            -- lets such a verdict normalize to (and classify) the same red. No native
+            -- job signal_key starts with 'coverage_'.
+            replaceRegexpOne(signal_key, '^coverage_', '') AS base_key,
             replaceRegexpOne(
                 replaceRegexpOne(
-                    signal_key,
+                    base_key,
                     ' \\[[^\\]]+\\]$', ''
                 ),
                 ', [0-9]+, [0-9]+, .+\\)', ')'
             ) AS adv_norm,
-            argMax(verdict, timestamp) AS verdict
+            -- Prefer a native (non-coverage_) verdict over a coverage_ one for the same
+            -- (commit, job); among equals, the latest timestamp wins.
+            argMax(verdict, (toUInt8(NOT startsWith(signal_key, 'coverage_')), timestamp)) AS verdict
         FROM misc.autorevert_advisor_verdicts
         WHERE
             repo = {repo: String}
             AND signal_source = 'job'
             AND timestamp >= {startTime: DateTime64(3)}
-        GROUP BY suspect_commit, signal_key
+        GROUP BY suspect_commit, base_key
     )
     GROUP BY head_sha, adv_norm
 ),

torchci/clickhouse_queries/flaky_trunk_runner_labels/query.sql

@@ -41,20 +41,27 @@ advisor_agg AS (
     FROM (
         SELECT
             toString(suspect_commit) AS head_sha,
+            -- Strip a leading "coverage_": coverage/backfill advisor dispatches use
+            -- signal_key = 'coverage_' + the native job signal_key, so stripping it
+            -- lets such a verdict normalize to (and classify) the same red. No native
+            -- job signal_key starts with 'coverage_'.
+            replaceRegexpOne(signal_key, '^coverage_', '') AS base_key,
             replaceRegexpOne(
                 replaceRegexpOne(
-                    signal_key,
+                    base_key,
                     ' \\[[^\\]]+\\]$', ''
                 ),
                 ', [0-9]+, [0-9]+, .+\\)', ')'
             ) AS adv_norm,
-            argMax(verdict, timestamp) AS verdict
+            -- Prefer a native (non-coverage_) verdict over a coverage_ one for the same
+            -- (commit, job); among equals, the latest timestamp wins.
+            argMax(verdict, (toUInt8(NOT startsWith(signal_key, 'coverage_')), timestamp)) AS verdict
         FROM misc.autorevert_advisor_verdicts
         WHERE
             repo = {repo: String}
             AND signal_source = 'job'
             AND timestamp >= {startTime: DateTime64(3)}
-        GROUP BY suspect_commit, signal_key
+        GROUP BY suspect_commit, base_key
     )
     GROUP BY head_sha, adv_norm
 ),

torchci/clickhouse_queries/flaky_trunk_timeseries/query.sql

@@ -55,20 +55,27 @@ advisor_agg AS (
     FROM (
         SELECT
             toString(suspect_commit) AS head_sha,
+            -- Strip a leading "coverage_": coverage/backfill advisor dispatches use
+            -- signal_key = 'coverage_' + the native job signal_key, so stripping it
+            -- lets such a verdict normalize to (and classify) the same red. No native
+            -- job signal_key starts with 'coverage_'.
+            replaceRegexpOne(signal_key, '^coverage_', '') AS base_key,
             replaceRegexpOne(
                 replaceRegexpOne(
-                    signal_key,
+                    base_key,
                     ' \\[[^\\]]+\\]$', ''
                 ),
                 ', [0-9]+, [0-9]+, .+\\)', ')'
             ) AS adv_norm,
-            argMax(verdict, timestamp) AS verdict
+            -- Prefer a native (non-coverage_) verdict over a coverage_ one for the same
+            -- (commit, job); among equals, the latest timestamp wins.
+            argMax(verdict, (toUInt8(NOT startsWith(signal_key, 'coverage_')), timestamp)) AS verdict
         FROM misc.autorevert_advisor_verdicts
         WHERE
             repo = {repo: String}
             AND signal_source = 'job'
             AND timestamp >= {startTime: DateTime64(3)}
-        GROUP BY suspect_commit, signal_key
+        GROUP BY suspect_commit, base_key
     )
     GROUP BY head_sha, adv_norm
 ),