9 posts tagged this topic.
AIDA's Settings AI Features form blocked Save with a "Select a model" error right after switching providers. The real defect sat two React commits away, inside a hidden native select Radix uses to bridge into native form submission — and it never raised an error of its own.
A Google Reviews listing-confirm route taught me that 'it saved' and 'everything it triggered afterward also worked' are two different claims, and a response body should never merge them into one.
Fixing a cross-tenant IDOR taught me that 'access denied' and 'does not exist' are the same message when the requester should never learn which one is true.
Two POS terminals sell the last unit in the same second. Inventory survives that only if every stock mutation goes through one function, one advisory lock, and one idempotency key — and if sub-cent rounding residue dies the moment stock hits zero.
AIDA's outbound email job derives References and In-Reply-To by querying every prior email-bearing message on the ticket. On a retry, the message being sent already has its own Message-ID persisted from the failed first attempt — without an explicit exclusion, it would show up in its own References chain.
Every webhook integration works in the demo. Production loses events in the silence after you reply 200 — so I built OmniSync around one guarantee: once an event is acknowledged, it is never silently lost. At-least-once delivery, idempotent everything.
AIDA's public intake route checks content length, per-file size, a running total, and byte-sniffed MIME type before saving an attachment. None of that mattered once Next.js's own proxy layer silently truncated the request body first, with no error back to the client.
Your scale ceiling is an arithmetic result, not a feeling — throughput is concurrency over latency, and a CPU-bound box tops out at cores divided by CPU-seconds per request. Most systems fall over at a small fraction of that number because of coordination and copies, not because the machine is small.
A model call is the slowest, priciest, least deterministic dependency in your whole system, so you ration it like an unreliable third-party API you call as rarely as you can. The practical skill is putting the model only where fuzzy understanding is genuinely required and doing the boring 80% with deterministic code you can test.