Pick an option on each card and press Done; your answer goes straight
back to the session. The raiser's pick is marked. "I'm not sure" is a real answer.
1
In plain terms. Our test-environment data mirror has been running about a day behind real production for at least the last day and a half, even after I gave it more headroom to catch up — should we stop restarting it on every unrelated code deploy, or make it remember its lookup data so a restart is cheap?
The full question, as raised
stage-replicator IteratorAge has held ~25.5-26.7h for 36+ continuous hours (since at least 2026-09-29 05:00 UTC), flapping the iterator-age alarm ~8 times without ever clearing. Lambda concurrency (peak ~64, baseline 4-8) stays far below the 150 reserved / 800 account limit; propflow-stage (on-demand) shows zero write throttles — neither is the bottleneck. Invocation duration spikes (up to 350s) line up exactly with deploy events — this Lambda's config gets touched on nearly every merge (~every 5-15min today), forcing cold containers to rebuild the in-memory PII dictionary (a known cost, partially mitigated 2 days ago by #13054's cold-start jitter). I bumped ParallelizationFactor 4->10 on the ESM as a safe, reversible stopgap, but that doesn't address why it never gets a sustained warm window to drain the backlog.
Why, and what was tried
The session raising this recommends: Skip redeploying it when its own code is unchanged Duration spikes track deploy timestamps exactly; a Lambda whose own code isn't in the diff shouldn't need a cold restart, and today's merge cadence already redeploys it every 5-15 minutes. Tried first, unsuccessfully: Pulled 36h of IteratorAge + alarm-flap history, correlated invocation-duration spikes to CloudTrail deploy events, checked Lambda concurrency vs reserved/account limits (not the bottleneck), checked propflow-stage write throttles (zero), bumped ParallelizationFactor 4->10 as a stopgap, and found the DLQ holds 428 parked batches (oldest ~2.5 days, RetryAttemptsExhausted) — a separate issue. Did not independently confirm the nightly batch's last successful run.