After getting burned by long-lived connections that slowly accumulate bad state (or get killed by the network) and then explode during peak traffic, I got strict about pg pooling. I keep the pool size small per instance and scale horizontally instead

With replicas, a user can create a record then immediately not see it due to replication lag. A simple mitigation is to pin subsequent reads to primary for a short window after writes.

Avoid writing rows when nothing changed—especially in batch jobs. Check changed? or compute the would-be update and skip if identical. This reduces bloat and autovacuum pressure.

Undo is a great UX improvement for destructive actions. My approach is: on destroy, soft-delete the record (or keep enough data to restore), remove it from the list via turbo_stream.remove, and append a toast with an “Undo” link. Clicking “Undo” hits

Tags are user input, so normalize aggressively: downcase, strip, collapse whitespace, remove duplicates. Doing this at write time keeps search/filter logic clean and avoids messy edge cases.

Production debugging is about logs that are searchable. Emit JSON with consistent keys (requestid, memberid, duration). Even if you later add Lograge, this structure maps cleanly.

Hotwire doesn’t replace caching—it makes it more valuable because you’re sending HTML frequently. I use fragment caching inside list partials and keep cache keys stable with cache blocks. The pattern is: cache each row by record, and cache the list wr

Spinners hide layout shifts and make an app feel slow even when it isn’t. Skeletons preserve layout and give users a sense of progress without jumping content around. I keep skeleton components simple and match the shape of the final UI. One practical

Sometimes you want to show non-blocking warnings (e.g., “link looks unreachable”) while still saving. Turbo streams can append warnings to a panel without rerendering the whole page.

Synthetic performance tests don’t always match what real users experience. Web Vitals reporting gives you field data (CLS, LCP, INP, etc.). I collect vitals on the client and POST them to a lightweight endpoint, then aggregate on the backend. Sampling

Not every part of a page needs to load immediately. For heavy panels (analytics charts, audit logs), I use a Turbo Frame with src and a tiny Stimulus controller that sets the src when the element becomes visible via IntersectionObserver. This avoids f

Search boxes are a classic place to accidentally fire a request on every keystroke. I debounce the value that triggers the query (not the input rendering) so typing stays responsive, and I cancel/ignore stale requests so results don’t arrive out of or