The cheap observability stack we wire up in an afternoon

The argument for doing it early

Observability is one of those line items that looks optional in the proposal and turns out to be load-bearing in week three. The first time a customer says 'something's broken' and you can't tell whether they mean the API, the auth flow, or their wifi — you wish you'd spent the afternoon.

The good news: you don't need a six-figure platform to cover 90% of what a small product needs. You need four boring pieces that talk to each other.

The four pieces

Error tracking. Sentry is still the default. Free tier is enough for most early-stage products. The win isn't catching errors — it's catching the ones users never bother to report. We turn on release tracking and source maps on day one so we know which deploy a regression came from.

Structured logs. Whatever your runtime, log JSON, not strings. Ship them somewhere queryable — Axiom, Logtail, even a Postgres table for a tiny project. The single biggest unlock is being able to ask 'show me every 500 from this user in the last hour' and get an answer in seconds.

One real-user metric. You don't need a full RUM suite. Pick the metric that matters for your product — time-to-interactive on the dashboard, time to first chart, whatever — and track it from real browsers. Vercel's Speed Insights or a five-line web-vitals snippet is enough.

A single status check. One synthetic ping that hits a critical path every minute. If your login flow is broken, you want to find out from a pager, not from a customer email at 9pm.

What we deliberately skip at this stage

• Distributed tracing — useful at scale, expensive at four endpoints
• Custom dashboards — premature until you know what to look at
• A/B test infrastructure — different problem, different tool
• Anything that requires its own on-call rota to operate

How long it actually takes

Sentry takes about twenty minutes including source maps. A logging pipeline takes an hour if you're using a managed service. Web vitals is a snippet. The synthetic check is a saved request in whatever monitor you like — five minutes.

Total: an afternoon, billed once. Compared to the cost of the first weekend you spend debugging a customer report blind, it pays for itself the first time something breaks.

The lesson

Observability isn't a maturity stage. It's a precondition for having a useful conversation about your product. Ship the four pieces above before your first real user logs in, and you'll spend the next six months reasoning about your software instead of guessing about it.