From 1caac95d2c60fdc98347d1fbc5179640255c2742 Mon Sep 17 00:00:00 2001
From: Seungpyo1007 <rush94434@gmail.com>
Date: Tue, 8 Sep 2026 13:35:33 +0900
Subject: [PATCH] fix(ci): stop weekly-refresh dying at the 6h ceiling before
 it dumps

Every run since at least 2026-08-03 has ended at 6:00:2x, always in
"Enrich benchmarks (all sources)". GitHub reports a job that hits its
6h ceiling as *cancelled* rather than failed, so nothing ever appeared
in the failed-run list while validate, integrity, the static dump and
the refresh PR were skipped every single week.

The cost is not theoretical: the published dump has drifted away from
data/ because the dump step never ran. 565 of 2,030 gpu pages, and 93
more across soc/watch/pda/brand, currently advertise a `verified` value
the record no longer has.

Give collection a 200-minute budget it checks between sources, so the
remainder of the pipeline always gets to run, and rotate the starting
point by ISO week so sources at the tail of the list are not starved
once the budget starts biting. Add a 330-minute job timeout so this
can never again be silently truncated by the platform ceiling.

Refs #1
---
 .github/workflows/weekly-refresh.yml | 33 +++++++++++++++++++++++++---
 1 file changed, 30 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/weekly-refresh.yml b/.github/workflows/weekly-refresh.yml
index 9f4d94d..c915f26 100644
--- a/.github/workflows/weekly-refresh.yml
+++ b/.github/workflows/weekly-refresh.yml
@@ -33,6 +33,9 @@ concurrency:
 jobs:
   refresh:
     runs-on: ubuntu-latest
+    # A job that runs into GitHub's 6h ceiling is reported as *cancelled*,
+    # not failed, so it never surfaces as a broken run. Stop short of it.
+    timeout-minutes: 330
     env:
       SLEEP: ${{ inputs.sleep || '1.0' }}
       TECHAPI_TOKEN: ${{ secrets.TECHAPI_TOKEN }}
@@ -80,11 +83,35 @@ jobs:
             fi
             echo "::endgroup::"
           }
-          for s in passmark cinebench-legacy cinebench-r23 cinebench-2024 \
-                   cinebench-nbc geekbench-nbc spec-cpu2006 topcpu-cpu; do
+          # The scrape grew past GitHub's 6h job ceiling, and an over-running
+          # job is reported as "cancelled" rather than failed — so nothing looked
+          # broken while validate, the static dump and the refresh PR were being
+          # skipped every week. Give collection a budget it cannot exceed and
+          # leave the remainder to the rest of the pipeline.
+          BUDGET_S=$(( 200 * 60 ))
+          DEADLINE=$(( $(date +%s) + BUDGET_S ))
+          budget_left() {
+            if [ "$(date +%s)" -ge "$DEADLINE" ]; then
+              echo "::warning::collection budget spent; remaining sources deferred to the next run"
+              return 1
+            fi
+          }
+          # Rotate where the list starts each week: with a fixed order the tail
+          # is starved forever once the budget begins to bite.
+          rotate() {
+            local -n _a="$1"
+            local n=${#_a[@]} off=$(( 10#$(date -u +%V) % ${#_a[@]} ))
+            printf '%s\n' "${_a[@]:off}" "${_a[@]:0:off}"
+          }
+          CPU_SOURCES=(passmark cinebench-legacy cinebench-r23 cinebench-2024
+                       cinebench-nbc geekbench-nbc spec-cpu2006 topcpu-cpu)
+          GPU_SOURCES=(blender timespy passmark-gpu topcpu-gpu)
+          for s in $(rotate CPU_SOURCES); do
+            budget_left || break
             run_enrich cpu "$s"
           done
-          for s in blender timespy passmark-gpu topcpu-gpu; do
+          for s in $(rotate GPU_SOURCES); do
+            budget_left || break
             run_enrich gpu "$s"
           done