Production systems, not demos
The gap between a working demo and a system your business can rely on is wider than most teams realise.
By Michele Sanna · Co-Founder & CPTO, SerSan
The demo is the easy part
A working demo on a senior engineer's laptop is one or two days of work. A system that handles a hundred thousand requests a day at p99 latency, fails gracefully, gets observed, gets rolled back, and obeys whatever regulatory regime sits over your industry. That is months.
Not every system needs all of that. A two-week automation that fails loudly and can be safely re-run is production-ready for what it does, and loading it with platform ceremony just makes it expensive. The list below is what production-grade means when production-grade is actually required.
What "production-ready" actually means
- Reproducible builds and infrastructure-as-code
- Observability: tracing, metrics, structured logs, model performance dashboards
- A defined rollback path that works under stress
- Cost monitoring with alerting before the bill becomes a problem
- Documentation that survives the engineer who wrote it leaving
The order of operations
In nine cases out of ten, the data plane needs work before the model is worth touching. Untangling the upstream pipelines, adding observability, and shoring up the boring infrastructure is usually 60–70% of the engagement.
Tags
Next