Performance
Performance capstone: free-recall synthesis
Retrieval beats re-reading. For each prompt, reconstruct a full answer from memory across the whole track before you open the model answer — the effort of pulling the connections together is what turns seven units into one discipline.
Reconstruct the track’s spine without looking back: the measurement loop, why Amdahl caps every speedup, the cost shapes you classify into, how to fix at the right layer, and why verify-then-enforce is what makes performance durable.
- 01State the five-step loop the whole track reduces to, and what each step prevents you from doing wrong.
- 02Why does Amdahl's law make 'profile first' non-negotiable, with a concrete number?
- 03Name the cost 'shapes' a hot path can take and the matching first move for each.
- 04A profile shows time in GC. Why is the fix usually not in the collector, and how does this generalise across the track?
- 05Explain the core tradeoff of batching and how the same per-op-cost reasoning shows up elsewhere in the track.
- 06What does it mean to verify and enforce a performance fix, and why is shipping a green local benchmark not enough?
If you could reconstruct each answer from memory, you hold the track’s spine: one loop — profile, classify, fix, verify, enforce — backed by Amdahl’s cap on optimising the wrong thing, a small vocabulary of cost shapes that map symptom to fix, the discipline of fixing at the cost’s own layer rather than a downstream knob, and the habit of proving the gain under real load and locking it in CI. Performance is the loop, not any one fix.