ci: invalidate poisoned shard-timing cache and guard future saves (#36568)

.github/workflows/server-ci.yml

@@ -247,6 +247,7 @@ jobs:
       artifact-pattern: postgres-server-test-logs-shard-*
       artifact-name: postgres-server-test-logs
       save-timing-cache: true
+      all-shards-passed: ${{ needs.test-postgres-normal.result == 'success' }}
 
   test-elasticsearch-v8:
     name: Elasticsearch v8 Compatibility

.github/workflows/server-test-merge-template.yml

@@ -16,6 +16,11 @@ on:
         required: false
         type: boolean
         default: false
+      all-shards-passed:
+        description: "Whether every upstream shard succeeded. Used to gate the timing-cache save so a single shard failure doesn't poison the cache with missing-package data."
+        required: false
+        type: boolean
+        default: false
 
 jobs:
   merge:
@@ -79,11 +84,17 @@ jobs:
             echo "has_timing=false" >> "$GITHUB_OUTPUT"
           fi
 
+      # Only save when every upstream shard succeeded. If even one shard
+      # failed/was killed, its gotestsum.json is missing and the merged report
+      # has no timings for that shard's packages — saving that would poison
+      # future shard splits (missing packages default to 1ms, all bin-pack
+      # onto the lightest shard, overloading it and repeating the failure).
       - name: Save test timing cache
-        if: inputs.save-timing-cache && steps.timing-prep.outputs.has_timing == 'true' && github.ref_name == github.event.repository.default_branch
+        if: inputs.save-timing-cache && inputs.all-shards-passed && steps.timing-prep.outputs.has_timing == 'true' && github.ref_name == github.event.repository.default_branch
         uses: actions/cache/save@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
         with:
           path: |
             server/prev-report.xml
             server/prev-gotestsum.json
-          key: server-test-timing-master-${{ github.run_id }}
+          # The v2 prefix matches the v2 restore prefix in server-test-template.yml.
+          key: server-test-timing-v2-master-${{ github.run_id }}

.github/workflows/server-test-template.yml

@@ -93,9 +93,15 @@ jobs:
             server/prev-gotestsum.json
           # Always restore from master — timing is only saved on the default
           # branch and is stable enough for shard balancing.
-          key: server-test-timing-master
+          # NOTE: the v2 prefix invalidates pre-existing caches that were
+          # poisoned by shard failures (a killed shard loses its gotestsum.json,
+          # so the merged report was missing those packages' timings; on the
+          # next run they all defaulted to 1ms and bin-packed onto the lightest
+          # shard, overloading it and perpetuating the cycle). See also the
+          # all-shards-passed guard in server-test-merge-template.yml.
+          key: server-test-timing-v2-master
           restore-keys: |
-            server-test-timing-
+            server-test-timing-v2-
 
       - name: Setup BUILD_IMAGE
         id: build