{
  "query": "histogram_quantile(0.95, checkout_request_seconds)",
  "window": "7d",
  "step": "5m"
}

{
  "before": 0.21,
  "after": 0.48,
  "changedAt": "2026-09-02T02:00:00Z"
}

{ "repo": "astralyxdev/checkout", "number": 4182 }

- const address = await lookup(line)\n+ const addresses = await lookupBatch(lines, { size: 200 })

PR #4182 batched the address lookup, and the address service started shedding load: its 429 rate went from 0.4% to 11.2% at 02:00 UTC — the same twenty minutes checkout p95 moved from 210ms to 480ms. The batch size in the diff is 200. The address service documents a limit of 50 per request and returns 429 above it without a Retry-After, so the client's backoff never engages and every batch costs a full timeout.