Search is the part of a product people reach for when they already know what they want. When it fails, it fails loudly, and it erodes trust faster than almost anything else.

Our original search was a thin wrapper around a database query. It was fine for a few hundred records and quietly miserable past a few thousand. Results were slow, ranking was arbitrary, and typos returned nothing at all.

The rewrite forced us to answer a harder question: what does a good result actually mean here? We learned that recency mattered more than we assumed, and that exact title matches should always win over fuzzy body matches.

The lesson was not technical. It was that the shape of your search reveals the shape of how people think about your product, and you cannot design one without understanding the other.

Why we rewrote our search from scratch

A healthy on-call rotation is one where most weeks are uneventful. The goal is not heroics at 3am. It is a system quiet enough that the person on call mostly sleeps.

When pages do come, the first question is never who made a mistake. It is what allowed a single mistake to reach production and wake someone up. Blameless review is not politeness, it is how you find the real fix.

We invest heavily in the runbook. A page that arrives with a clear, tested set of steps is a minor inconvenience. The same page without one is a stressful archaeology dig through unfamiliar systems.

Over time the measure of a good on-call culture is simple. The rotation gets quieter, the runbooks get shorter, and nobody dreads the week their name comes up.

Engineering

Why we rewrote our search from scratch

What a good on-call week looks like