Performance optimization

Performance Optimization

Make slow things fast without guessing. Performance work that is not driven by a measurement is decoration. The cross-cutting principles hub is the coding-standards skill; this skill is the optimisation detail.

---

When to activate

• An application or endpoint is slow, or a deployment regressed a latency or throughput metric.
• You need a baseline before optimising, or want to verify an optimisation actually moved the metric.
• Reviewing a change for an obvious performance trap (a query in a loop, an unbounded fetch, a sync call that should run concurrently).

---

Measure first

• Do not optimise before measuring. Profile, find the one real bottleneck, then fix that one thing. The bottleneck is almost never where intuition points.
• Set a target before you start (a latency budget, a throughput number) so you know when you are done, and measure again after to confirm the change moved the metric and did not just move the cost elsewhere.

---

Data access

• Avoid the N+1 query problem. Batch related lookups, and never run a query inside a loop when one query would do.
• Fetch only the data you need: select only the columns a query uses, and page large result sets rather than loading everything.
• Add the indexes the real queries require, and confirm with the database's own query plan that they are used. An index nobody queries is write-cost for no read benefit.

---

Caching

Cache where reads dominate and the data tolerates slight staleness, with a clear rule for invalidating it. A cache without an invalidation story is a correctness bug waiting to surface. Name the staleness you accept and the event that clears the entry.

---

Memory and concurrency

• Build a string from many pieces with a join or a buffer, never by repeated concatenation in a loop.
• Stream large data lazily instead of loading it all into memory. A file or result set that fits today overflows tomorrow.
• Run independent input and output work concurrently rather than one call after another. Sequential awaits of unrelated calls waste the slowest one's whole duration.