perf(run-engine): dequeue gauges sample once, at return

The gauge slot is single-valued and the last write wins, so on the
success path the post-admission resample made the entry sample pure
waste. A return wrapper computes the gauge exactly once per call at
exit, keeping every emitted value identical while dropping the
discarded reads (about six per plain dequeue, ten per keyed dequeue
at full sampling).

Author: matt-aitken

internal-packages/run-engine/src/run-queue/index.ts

@@ -173,8 +173,9 @@ local __qm_g = false
 local function __qmret(r) if r == nil then r = false end return {r, __qm_g} end`;
 
 // Fresh-read gauge for splice points with no reusable locals: enqueue slow-path (before
-// return 0) and the base dequeue top. Gated on the last ARGV so it is inert unless the
-// caller opts in. CK queues emit per-subqueue depth (queue_name aggregates via the MV).
+// return 0) and the base dequeue's sample-at-return wrapper. Gated on the last ARGV so it
+// is inert unless the caller opts in. CK queues emit per-subqueue depth (queue_name
+// aggregates via the MV).
 const QUEUE_METRICS_GAUGE_LUA = createMetricsGaugeComputeLua({
   enabledArg: "ARGV[#ARGV] == '1'",
   queued: "redis.call('ZCARD', queueKey)",
@@ -4696,7 +4697,16 @@ local gatesEnabled = ARGV[7] == '1'
 local totalConcurrencyEnabled = ARGV[8] == '1'
 ${QUEUE_METRICS_GAUGE_PRELUDE}
 ${QUEUE_GATES_LUA_HELPERS}
+-- Sample-at-return: the gauge is computed once, by the return wrapper, so every
+-- exit emits the state as of that exit (post-admission on the success path) and
+-- no path pays for a sample that a later one would overwrite.
+local function __qmsample()
 ${QUEUE_METRICS_GAUGE_LUA}
+end
+do
+  local __qmret_inner = __qmret
+  __qmret = function(r) __qmsample() return __qmret_inner(r) end
+end
 
 -- Check current env concurrency against the limit
 local envCurrentConcurrency = tonumber(redis.call('SCARD', envCurrentConcurrencyKey) or '0')
@@ -4808,10 +4818,6 @@ else
   redis.call('ZADD', masterQueueKey, earliestMessage[2], queueName)
 end
 
--- Re-sample the gauge so the emitted snapshot includes this batch's admissions;
--- the top-of-script sample only covers the early returns where nothing was admitted.
-${QUEUE_METRICS_GAUGE_LUA}
-
 -- Return results as a flat array: [messageId1, messageScore1, messagePayload1, messageId2, messageScore2, messagePayload2, ...]
 return __qmret(results)
       `,
@@ -5005,7 +5011,16 @@ local totalConcurrencyEnabled = ARGV[7] == '1'
 local gatesEnabled = ARGV[8] == '1'
 ${QUEUE_METRICS_GAUGE_PRELUDE}
 ${QUEUE_GATES_LUA_HELPERS}
+-- Sample-at-return: the gauge is computed once, by the return wrapper, so every
+-- exit emits the state as of that exit (post-admission on the success path) and
+-- no path pays for a sample that a later one would overwrite.
+local function __qmsample()
 ${QUEUE_METRICS_CK_DEQUEUE_GAUGE_LUA}
+end
+do
+  local __qmret_inner = __qmret
+  __qmret = function(r) __qmsample() return __qmret_inner(r) end
+end
 
 local function decrLengthCounter()
   if tonumber(redis.call('GET', lengthCounterKey) or '0') > 0 then
@@ -5194,10 +5209,6 @@ else
   redis.call('ZADD', masterQueueKey, earliestIdx[2], ckWildcardName)
 end
 
--- Re-sample the gauge so the emitted snapshot includes this batch's admissions;
--- the top-of-script sample only covers the early returns where nothing was admitted.
-${QUEUE_METRICS_CK_DEQUEUE_GAUGE_LUA}
-
 return __qmret(results)
       `,
     });