BUILD-11593 Render CPU request as the avg denominator (schema 3)

ci-metrics/job-completed.sh

@@ -144,8 +144,18 @@ else
         'BEGIN{ if (p+0>0) printf "%.3f", q/p; }')
 fi
 
-# Online CPU count — the denominator for utilisation when there is no cgroup quota
-# (the common case for our runners). Overridable for deterministic tests.
+# CPU-avg denominator candidates, in preference order:
+#   1. cpu_limit_cores  — hard cgroup quota (above), when set.
+#   2. cpu_request_cores — the runner's CPU request from the Downward API env var
+#      CI_METRICS_CPU_REQUEST_MILLI (millicores). Exact; the right denominator for our
+#      burstable ARC runners (no quota, request set). Absent on WarpBuild.
+#   3. cpu_online_count — nproc. Only correct when the runner owns the host (WarpBuild
+#      dedicated VM); on ARC this is the whole node, so it's the last resort.
+if [[ "${CI_METRICS_CPU_REQUEST_MILLI:-}" =~ ^[0-9]+$ ]] && (( CI_METRICS_CPU_REQUEST_MILLI > 0 )); then
+    cpu_request_cores=$(awk -v m="$CI_METRICS_CPU_REQUEST_MILLI" 'BEGIN{printf "%.3f", m/1000}')
+else
+    cpu_request_cores=""
+fi
 cpu_online_count="${CI_METRICS_NPROC:-$(nproc 2>/dev/null || true)}"
 [[ "$cpu_online_count" =~ ^[0-9]+$ ]] || cpu_online_count=""
 
@@ -298,7 +308,7 @@ for iface in "${!net_rx[@]}"; do
 done
 
 job_metrics_json=$(cat <<EOF
-{"schema_version":2,"captured_at":"$captured_at","duration_seconds":$(num "$duration_seconds"),"cgroup":{"version":2,"cpu":{"usage_seconds":$(num "$cpu_usage_seconds"),"throttled_seconds":$(num "$cpu_throttled_seconds"),"nr_throttled":$(num "$cpu_nr_throttled"),"limit_cores":$(num "$cpu_limit_cores"),"online_count":$(num "$cpu_online_count"),"avg_utilization":$(num "$cpu_avg_utilization"),"throttle_rate":$(num "$cpu_throttle_rate"),"pressure_some_avg10":$(num "$pressure_some_avg10"),"pressure_some_avg60":$(num "$pressure_some_avg60"),"pressure_some_avg300":$(num "$pressure_some_avg300")},"memory":{"peak_bytes":$(num "$memory_peak_bytes"),"limit_bytes":$(num "$memory_limit_bytes"),"peak_utilization":$(num "$memory_peak_utilization"),"oom":$(num "$memory_oom"),"oom_kill":$(num "$memory_oom_kill")}},"net":{"rx_bytes":${net_rx_total},"tx_bytes":${net_tx_total},"by_interface":{${json_by_interface}}},"disk":{"path":${disk_path_json},"total_bytes":$(num "$disk_total_bytes"),"used_bytes":$(num "$disk_used_bytes"),"available_bytes":$(num "$disk_avail_bytes"),"utilization":$(num "$disk_utilization")}}
+{"schema_version":3,"captured_at":"$captured_at","duration_seconds":$(num "$duration_seconds"),"cgroup":{"version":2,"cpu":{"usage_seconds":$(num "$cpu_usage_seconds"),"throttled_seconds":$(num "$cpu_throttled_seconds"),"nr_throttled":$(num "$cpu_nr_throttled"),"limit_cores":$(num "$cpu_limit_cores"),"request_cores":$(num "$cpu_request_cores"),"online_count":$(num "$cpu_online_count"),"avg_utilization":$(num "$cpu_avg_utilization"),"throttle_rate":$(num "$cpu_throttle_rate"),"pressure_some_avg10":$(num "$pressure_some_avg10"),"pressure_some_avg60":$(num "$pressure_some_avg60"),"pressure_some_avg300":$(num "$pressure_some_avg300")},"memory":{"peak_bytes":$(num "$memory_peak_bytes"),"limit_bytes":$(num "$memory_limit_bytes"),"peak_utilization":$(num "$memory_peak_utilization"),"oom":$(num "$memory_oom"),"oom_kill":$(num "$memory_oom_kill")}},"net":{"rx_bytes":${net_rx_total},"tx_bytes":${net_tx_total},"by_interface":{${json_by_interface}}},"disk":{"path":${disk_path_json},"total_bytes":$(num "$disk_total_bytes"),"used_bytes":$(num "$disk_used_bytes"),"available_bytes":$(num "$disk_avail_bytes"),"utilization":$(num "$disk_utilization")}}
 EOF
 )
 
@@ -351,21 +361,31 @@ summary_target="${GITHUB_STEP_SUMMARY:-/dev/null}"
 
 # Pretty values
 
-# CPU avg: mean cores used over the job, shown against the cores available so the number
-# is actionable for right-sizing. Denominator is the cgroup quota when one is set,
-# otherwise the online CPU count. Falls back to bare cores, then n/a.
+# CPU avg: mean cores used over the job, shown against a denominator so the number is
+# actionable for right-sizing. Denominator preference:
+#   1. cgroup quota  -> "/ N cores (P%)"            (hard limit)
+#   2. CPU request   -> "/ N cores requested (P%)"  (ARC burstable; P can exceed 100% when bursting)
+#   3. nproc         -> "/ N cores available (P%)"  (WarpBuild dedicated VM)
+# Falls back to bare cores, then n/a.
+# This rendering mirrors _rci_cpu_cell in report-ci-metrics/lib.sh; the expected output per tier
+# is pinned by the "_rci_cpu_cell() denominator preference" spec — keep both in sync if changed.
 cpu_avg_cores=""
 if [[ -n "$cpu_usage_seconds" && -n "$duration_seconds" ]]; then
     cpu_avg_cores=$(awk -v u="$cpu_usage_seconds" -v d="$duration_seconds" \
         'BEGIN{ if (d>0) printf "%.2f", u/d }')
 fi
 if [[ -n "$cpu_avg_cores" && -n "$cpu_limit_cores" && -n "$cpu_avg_utilization" ]]; then
     v_cpu_avg="${cpu_avg_cores} / $(round2 "$cpu_limit_cores") cores ($(pct0 "$cpu_avg_utilization")%)"
+elif [[ -n "$cpu_avg_cores" && -n "$cpu_request_cores" ]] \
+     && awk -v r="$cpu_request_cores" 'BEGIN{exit !(r+0>0)}'; then
+    cpu_avg_pct=$(awk -v c="$cpu_avg_cores" -v r="$cpu_request_cores" \
+        'BEGIN{ printf "%.0f", (c/r)*100 }')
+    v_cpu_avg="${cpu_avg_cores} / $(round2 "$cpu_request_cores") cores requested (${cpu_avg_pct}%)"
 elif [[ -n "$cpu_avg_cores" && -n "$cpu_online_count" ]] \
      && (( cpu_online_count > 0 )); then
     cpu_avg_pct=$(awk -v c="$cpu_avg_cores" -v n="$cpu_online_count" \
         'BEGIN{ printf "%.0f", (c/n)*100 }')
-    v_cpu_avg="${cpu_avg_cores} / ${cpu_online_count} cores (${cpu_avg_pct}%)"
+    v_cpu_avg="${cpu_avg_cores} / ${cpu_online_count} cores available (${cpu_avg_pct}%)"
 elif [[ -n "$cpu_avg_cores" ]]; then
     v_cpu_avg="${cpu_avg_cores} cores"
 else

report-ci-metrics/lib.sh

@@ -56,13 +56,12 @@ _rci_fmt_bytes() {
   fi
 }
 
-# CPU-avg display cell for one job's JSON, mirroring the hook's step-summary logic:
-#   cores = usage_seconds / duration_seconds (2dp);
-#   denominator = limit_cores when avg_utilization is known, else online_count;
-#   "<cores> / <avail> cores (<pct>%)", falling back to bare cores, then "n/a".
+# CPU-avg display cell for one job's JSON, mirroring the hook's step-summary logic.
+# Denominator preference: limit_cores -> request_cores ("requested", ARC burstable) ->
+# online_count ("available", WarpBuild VM) -> bare cores -> "n/a".
 _rci_cpu_cell() {
   local json=$1
-  jq -r '.cgroup.cpu as $c | .duration_seconds as $d | if ($c.usage_seconds != null and $d != null and $d > 0) then (($c.usage_seconds / $d) * 100 | round / 100) as $cores | if ($c.limit_cores != null and $c.avg_utilization != null) then "\($cores) / \(($c.limit_cores*100|round)/100) cores (\(($c.avg_utilization*100)|round)%)" elif ($c.online_count != null and $c.online_count > 0) then "\($cores) / \($c.online_count) cores (\((($cores/$c.online_count)*100)|round)%)" else "\($cores) cores" end else "n/a" end' <<< "$json"
+  jq -r '.cgroup.cpu as $c | .duration_seconds as $d | if ($c.usage_seconds != null and $d != null and $d > 0) then (($c.usage_seconds / $d) * 100 | round / 100) as $cores | if ($c.limit_cores != null and $c.avg_utilization != null) then "\($cores) / \(($c.limit_cores*100|round)/100) cores (\(($c.avg_utilization*100)|round)%)" elif ($c.request_cores != null and $c.request_cores > 0) then "\($cores) / \(($c.request_cores*100|round)/100) cores requested (\((($cores/$c.request_cores)*100)|round)%)" elif ($c.online_count != null and $c.online_count > 0) then "\($cores) / \($c.online_count) cores available (\((($cores/$c.online_count)*100)|round)%)" else "\($cores) cores" end else "n/a" end' <<< "$json"
 }
 
 # Sum a numeric jq path across all record JSONs; nulls count as 0. Echoes an integer-ish sum.

spec/report-ci-metrics_spec.sh

@@ -181,11 +181,11 @@ Describe 'report-ci-metrics/lib.sh'
     End
   End
 
-  # schema_version 2 fixtures (one "<name>\t<json>" record each):
+  # schema_version 3 fixtures (one "<name>\t<json>" record each):
   #   build = cache restored+saved, no flags; test = OOM-killed; lint = no disk, throttled.
-  J_BUILD='{"schema_version":2,"captured_at":"t","duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":40.0,"throttled_seconds":0.0,"nr_throttled":0,"limit_cores":2.0,"online_count":4,"avg_utilization":0.32,"throttle_rate":null,"pressure_some_avg10":null,"pressure_some_avg60":null,"pressure_some_avg300":null},"memory":{"peak_bytes":3435973836,"limit_bytes":4294967296,"peak_utilization":0.80,"oom":0,"oom_kill":0}},"net":{"rx_bytes":1073741824,"tx_bytes":209715200,"by_interface":{}},"disk":{"path":"/","total_bytes":32212254720,"used_bytes":6442450944,"available_bytes":null,"utilization":0.20},"cache":[{"key":"maven-abc","cache_hit":true,"restore_key_hit":null,"backend":"s3","size_bytes_restored":471859200,"saved":true,"size_bytes_at_end":492832000}]}'
-  J_TEST='{"schema_version":2,"captured_at":"t","duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":10.0,"throttled_seconds":0.0,"nr_throttled":0,"limit_cores":null,"online_count":4,"avg_utilization":null,"throttle_rate":null,"pressure_some_avg10":null,"pressure_some_avg60":null,"pressure_some_avg300":null},"memory":{"peak_bytes":104857600,"limit_bytes":4294967296,"peak_utilization":0.02,"oom":0,"oom_kill":1}},"net":{"rx_bytes":524288000,"tx_bytes":104857600,"by_interface":{}},"disk":{"path":"/","total_bytes":32212254720,"used_bytes":3221225472,"available_bytes":null,"utilization":0.10},"cache":[{"key":"npm-xyz","cache_hit":true,"restore_key_hit":null,"backend":"gha","size_bytes_restored":104857600,"saved":false,"size_bytes_at_end":null}]}'
-  J_LINT='{"schema_version":2,"captured_at":"t","duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":2.0,"throttled_seconds":3.5,"nr_throttled":7,"limit_cores":2.0,"online_count":4,"avg_utilization":0.01,"throttle_rate":0.05,"pressure_some_avg10":null,"pressure_some_avg60":null,"pressure_some_avg300":null},"memory":{"peak_bytes":52428800,"limit_bytes":4294967296,"peak_utilization":0.01,"oom":0,"oom_kill":0}},"net":{"rx_bytes":0,"tx_bytes":0,"by_interface":{}},"disk":{"path":"/","total_bytes":null,"used_bytes":null,"available_bytes":null,"utilization":null}}'
+  J_BUILD='{"schema_version":3,"captured_at":"t","duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":40.0,"throttled_seconds":0.0,"nr_throttled":0,"limit_cores":2.0,"request_cores":1.0,"online_count":4,"avg_utilization":0.32,"throttle_rate":null,"pressure_some_avg10":null,"pressure_some_avg60":null,"pressure_some_avg300":null},"memory":{"peak_bytes":3435973836,"limit_bytes":4294967296,"peak_utilization":0.80,"oom":0,"oom_kill":0}},"net":{"rx_bytes":1073741824,"tx_bytes":209715200,"by_interface":{}},"disk":{"path":"/","total_bytes":32212254720,"used_bytes":6442450944,"available_bytes":null,"utilization":0.20},"cache":[{"key":"maven-abc","cache_hit":true,"restore_key_hit":null,"backend":"s3","size_bytes_restored":471859200,"saved":true,"size_bytes_at_end":492832000}]}'
+  J_TEST='{"schema_version":3,"captured_at":"t","duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":10.0,"throttled_seconds":0.0,"nr_throttled":0,"limit_cores":null,"request_cores":1.0,"online_count":4,"avg_utilization":null,"throttle_rate":null,"pressure_some_avg10":null,"pressure_some_avg60":null,"pressure_some_avg300":null},"memory":{"peak_bytes":104857600,"limit_bytes":4294967296,"peak_utilization":0.02,"oom":0,"oom_kill":1}},"net":{"rx_bytes":524288000,"tx_bytes":104857600,"by_interface":{}},"disk":{"path":"/","total_bytes":32212254720,"used_bytes":3221225472,"available_bytes":null,"utilization":0.10},"cache":[{"key":"npm-xyz","cache_hit":true,"restore_key_hit":null,"backend":"gha","size_bytes_restored":104857600,"saved":false,"size_bytes_at_end":null}]}'
+  J_LINT='{"schema_version":3,"captured_at":"t","duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":2.0,"throttled_seconds":3.5,"nr_throttled":7,"limit_cores":2.0,"request_cores":1.0,"online_count":4,"avg_utilization":0.01,"throttle_rate":0.05,"pressure_some_avg10":null,"pressure_some_avg60":null,"pressure_some_avg300":null},"memory":{"peak_bytes":52428800,"limit_bytes":4294967296,"peak_utilization":0.01,"oom":0,"oom_kill":0}},"net":{"rx_bytes":0,"tx_bytes":0,"by_interface":{}},"disk":{"path":"/","total_bytes":null,"used_bytes":null,"available_bytes":null,"utilization":null}}'
 
   Describe 'render_headline()'
     It 'renders correct totals: CPU-seconds, worst peak mem with job, net, cache'
@@ -289,6 +289,33 @@ Describe 'report-ci-metrics/lib.sh'
     End
   End
 
+  Describe '_rci_cpu_cell() denominator preference'
+    # cores = usage/duration. Denominator: limit -> request -> online_count -> bare.
+    It 'uses the cgroup limit when present'
+      j='{"duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":31.0,"limit_cores":2.0,"avg_utilization":0.25,"request_cores":1.0,"online_count":32}}}'
+      When call _rci_cpu_cell "$j"
+      The output should equal '0.5 / 2 cores (25%)'
+    End
+
+    It 'uses the CPU request (not nproc) when there is no limit — the ARC bug fix'
+      j='{"duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":4.34,"limit_cores":null,"avg_utilization":null,"request_cores":1.0,"online_count":32}}}'
+      When call _rci_cpu_cell "$j"
+      The output should equal '0.07 / 1 cores requested (7%)'
+    End
+
+    It 'falls back to online_count (available) on WarpBuild when no limit/request'
+      j='{"duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":31.0,"limit_cores":null,"avg_utilization":null,"request_cores":null,"online_count":8}}}'
+      When call _rci_cpu_cell "$j"
+      The output should equal '0.5 / 8 cores available (6%)'
+    End
+
+    It 'shows bare cores when no denominator is available'
+      j='{"duration_seconds":62.0,"cgroup":{"cpu":{"usage_seconds":31.0,"limit_cores":null,"avg_utilization":null,"request_cores":null,"online_count":null}}}'
+      When call _rci_cpu_cell "$j"
+      The output should equal '0.5 cores'
+    End
+  End
+
   Describe 'render_cache_fold()'
     It 'renders a cache block when at least one job has a cache entry'
       records=$(printf '%s\t%s\n' 'build' "$J_BUILD" 'test' "$J_TEST" 'lint' "$J_LINT")