While auditing the spec test-suites I noticed there's no benchmark exercising the blocked → unblocked client path. None of the 380 memtier_benchmark-*.yml files use BLPOP, BRPOP, BLMOVE, BLMPOP, BZPOPMIN, BZMPOP, or BLOCK on XREAD/XREADGROUP, so regressions in blocked.c (blockForKeys, processUnblockedClients, unblockClient, serveClientBlockedOnList) go undetected by CE Performance.
This is performance-sensitive because:
- Blocking clients live on the per-key blocked-on-keys index; unblock latency directly drives p99 for pub-sub-like workflows on lists and streams (work queues, fan-out notifications).
processUnblockedClients is iterated in beforeSleep() every event-loop tick — bookkeeping cost there scales with the unblocked queue length.signalKeyAsReady / handleClientsBlockedOnKeys walk the ready-key list and pop one blocked client at a time; their cost shows up under high churn.
Suggested initial coverage:
memtier_benchmark-Nclients-blpop-1key-rotated-pipeline-1.yml — 1 producer doing LPUSH ... <ts>, N consumers BLPOP key 0, measure rate and p50/p99 unblock latency.memtier_benchmark-Nclients-bzpopmin-1key-rotated-pipeline-1.yml — same shape for sorted sets.memtier_benchmark-Nclients-xread-block-consumer-group.yml — XADD producer + N XREADGROUP ... BLOCK 0 consumers.
Happy to take this on if it's helpful — let me know if there's a preferred shape (e.g. a single rotating-key playbook vs. per-command files).
While auditing the spec test-suites I noticed there's no benchmark exercising the blocked → unblocked client path. None of the 380
memtier_benchmark-*.ymlfiles use BLPOP, BRPOP, BLMOVE, BLMPOP, BZPOPMIN, BZMPOP, or BLOCK on XREAD/XREADGROUP, so regressions inblocked.c(blockForKeys,processUnblockedClients,unblockClient,serveClientBlockedOnList) go undetected by CE Performance.This is performance-sensitive because:
processUnblockedClientsis iterated inbeforeSleep()every event-loop tick — bookkeeping cost there scales with the unblocked queue length.signalKeyAsReady/handleClientsBlockedOnKeyswalk the ready-key list and pop one blocked client at a time; their cost shows up under high churn.Suggested initial coverage:
memtier_benchmark-Nclients-blpop-1key-rotated-pipeline-1.yml— 1 producer doingLPUSH ... <ts>, N consumersBLPOP key 0, measure rate and p50/p99 unblock latency.memtier_benchmark-Nclients-bzpopmin-1key-rotated-pipeline-1.yml— same shape for sorted sets.memtier_benchmark-Nclients-xread-block-consumer-group.yml— XADD producer + NXREADGROUP ... BLOCK 0consumers.Happy to take this on if it's helpful — let me know if there's a preferred shape (e.g. a single rotating-key playbook vs. per-command files).