Revisiting Search Infrastructure After 90 Sprint in Production

Database connection pooling was our biggest blind spot. Under normal load, direct connections worked fine. But during traffic spikes, the database would hit its connection limit and cascade failures across all services. A simple PgBouncer setup eliminated the issue entirely.

We started this project with a clear hypothesis: the existing approach was costing us more in maintenance time than the migration would cost upfront. Three months later, the data confirmed we were right — but the journey was far bumpier than expected.

Error handling deserves as much design attention as the happy path. We created a taxonomy of error types — retryable, user-fixable, operator-fixable, and fatal — and built standard handling patterns for each. Support tickets dropped by half because users finally got actionable error messages instead of generic 500 pages.

Authentication turned out to be the most politically charged decision in the entire project. Every team had opinions about OAuth providers, session management strategies, and token lifetimes. We eventually settled on a pragmatic middle ground that nobody loved but everyone could live with.

Feature flags transformed our release process more than any CI/CD improvement. Decoupling deployment from release meant we could merge code daily, test in production with internal users, and gradually roll out to customers — all while maintaining the ability to instantly revert without a code deployment.

Our initial benchmark numbers looked promising in staging but fell apart under production traffic patterns. The difference? Staging used uniform request distributions while real users exhibit bursty, correlated behavior that exposes different bottlenecks entirely.

What Changed

Caching is deceptively simple in concept and endlessly complex in practice. Our first implementation had cache stampede issues under load, our second had stale data bugs that took weeks to diagnose, and our third attempt finally got it right by using a combination of TTLs, background refresh, and circuit breakers.

Synthetic monitoring catches problems that real-user monitoring misses: slow third-party scripts, broken OAuth flows at 3 AM, and regional CDN issues. We run synthetic checks from twelve global locations every five minutes and page the on-call engineer if any critical path degrades beyond thresholds.

What worked for us won’t work for everyone. Context matters enormously. But we hope sharing our experience saves someone else from repeating our more expensive mistakes.