You are a performance-driven backend engineer using Claude to ensure APIs remain fast, efficient, and resilient as traffic grows.
🚀 High-Performance API Design
- Identify endpoints that become bottlenecks under peak load
- Avoid expensive synchronous work inside request/response cycles
- Consider pagination and streaming for large result sets
Reference: https://fastapi.tiangolo.com/async/
⚡ Async-First Concurrency
- Prefer async I/O where latency matters most
- Audit blocking calls that freeze event loops
- Separate CPU-heavy tasks into background workers
Reference: https://docs.python.org/3/library/asyncio.html
🧮 Efficient Data Access
- Use index coverage and reduce redundant round-trips
- Cache stable reads using Redis or app-level memory where safe
- Pre-compute expensive joins and summaries where helpful
Reference: https://docs.sqlalchemy.org/
📦 Smart Caching & Content Strategy
- Cache whole responses when data updates infrequently
- Maintain cache-invalidations as part of business logic
- Avoid caching personalized or highly sensitive payloads
Reference: https://developer.mozilla.org/docs/Web/HTTP/Caching
📉 Load Patterns & Observability
- Track error rates, queuing time, and dependency latency
- Analyze p95 & p99 metrics to understand real bottlenecks
- Collect distributed traces for cross-service debugging
Reference: https://opentelemetry.io/
🔁 Performance Regression Prevention
- Performance tests run in CI before new releases
- Claude summarizes profiling results in plain English
- Deprioritize "tiny wins" — fix bottlenecks that matter
- Claude is powerful at recommending impact-based prioritization
🧱 Infrastructure Scaling Safety
- Autoscale based on user-experience signals, not raw CPU
- Split workloads so failures degrade gracefully, not catastrophically
- Trust slow-roll deployments to limit blast radius
🔑 Core Performance & Scale Principles
- Data access and serialization dominate latency
- Async isn't magic — measure first
- Observability is part of performance
- Claude helps you explain and fix "why slow?"
- Optimization is never one-and-done