| name | performance-engineering |
| description | Use for latency, throughput, CPU, memory, allocations, large datasets, hot loops, serialization, database performance, concurrency, backpressure, caching, or capacity planning. |
Performance Engineering Skill
Workflow
- Define workload and budget: rate, concurrency, payload/data size, latency percentiles, throughput, CPU, memory, and error tolerance.
- Collect a representative baseline using existing benchmarks, profiler, tracing, database plans, or load tooling.
- Decompose end-to-end cost into compute, allocation/copy, storage, network, contention, queueing, and retry components.
- Rank bottlenecks by measured contribution.
- Choose the simplest change with the highest expected impact.
- For hot paths, assess streaming, zero/bounded copy, batching, preallocation, contiguous layout, cache locality, and bounded concurrency.
- Preserve correctness with a reference/property test and add a performance regression threshold.
- Re-run measurements under the same conditions and report variance.
Complexity annotation
For each performance-critical function, document:
- time: best/typical/worst when materially different;
- auxiliary space;
- dominant allocations and copies;
- I/O and synchronization operations;
- input characteristic that drives the worst case.
Stop conditions
Stop and request explicit design review before adding unsafe memory, lock-free algorithms, custom allocators, unbounded caches, or weaker durability/security semantics.
