Shipping a fix to someone else's repo

One of our portfolio apps was quietly losing orders. Not errors — delays. The billing webhooks were processed serially, and after their summer growth, the queue was measured in days while the dashboard showed a green checkmark.

The bug was the architecture, not the code: one background worker, one queue, all event types lumped together. Restore-progress looked fine; checkout did not.

The fix, in three commits:

• Split the queue by event type — checkout and cancellation get their own lanes with their own SLAs. Two hours.
• Add a dead-letter queue with alerting on retry depth instead of failure count. One afternoon.
• Backfill and replay from the provider’s event log, then a two-week shadow mode where both pipelines ran and pivoted to the new one. This was the boring, load-bearing part.

Total: a long weekend of our engineers’ time, no customer-visible incident. The founders did the product work; the plumbing came from us.

That’s the whole deal of capital + code. The check is table stakes; the checkout lane is ours to unblock.