---
description: Performance budgets, reliability design, caching, and scalability safeguards
alwaysApply: true
---

# Performance & Reliability

Performance and reliability must be designed, measured, and protected. Do not optimize blindly or defer obvious scalability risks.

## Budgets
- Define expected latency, throughput, payload size, memory, storage, and concurrency for high-impact flows.
- User-facing flows should consider p95 and p99 latency, not only averages.
- Establish performance budgets for page weight, render time, API latency, and query count when applicable.
- If no budget exists, propose one before making risky design choices.

## Hot Paths
- Avoid N+1 queries, unbounded loops with I/O, repeated serialization, and large synchronous work on request paths.
- Paginate or stream large datasets. Do not load unbounded result sets into memory.
- Load the `collection-standards` skill when implementation, QA automation, or DevOps scripting processes repeated collections, joins lists, scans logs, builds command matrices, or must prove O(n) or bounded complexity.
- Keep expensive work outside user-facing request paths through queues, jobs, or precomputation.
- Measure before and after performance changes and report the evidence.

## Caching
- Cache only with clear ownership, TTL, invalidation strategy, and stale-data behavior.
- Do not cache sensitive data unless storage, access, and expiration controls are explicit.
- Prefer simple, observable cache strategies before multi-layer cache complexity.
- Document cache keys and invalidation triggers for shared caches.

## Reliability
- Design for graceful degradation when dependencies fail.
- Add timeouts, retries, backoff, and circuit breakers for networked dependencies.
- Use bulkheads or queues to protect critical paths from non-critical workload spikes.
- Reliability-sensitive changes need load, stress, or synthetic tests when practical.