# Performance Lens > Performance review lens for evaluating algorithmic efficiency, resource usage, and concurrency efficiency. Used by review orchestrators — not invoked directly. - Skill: `atomicinnovation/performance-lens` (Agent Skill) - Install (CLI): `npx skillmds@latest add atomicinnovation/performance-lens` - Raw SKILL.md: https://api.skillmd.com/api/skills/atomicinnovation/performance-lens/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: atomicinnovation (https://skillmd.com/u/atomicinnovation) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/atomicinnovation/performance-lens --- # Performance Lens Review as a capacity planner identifying where the system will bottleneck under load. ## Core Responsibilities 1. **Evaluate Algorithmic Efficiency and Data Structure Selection** - Assess time and space complexity of algorithms and data structures - Identify unnecessary iteration (nested loops, redundant passes, repeated lookups) - Check data structure fitness — maps vs lists for lookups, sets vs arrays for membership, appropriate use of indexes - Evaluate sorting and searching strategies for dataset sizes involved - Identify opportunities to reduce work (early returns, short-circuiting, memoisation) 2. **Assess Resource Efficiency and I/O Performance** - Check for memory allocation patterns in hot paths (object creation in loops, unbounded caches, string concatenation in loops) - Evaluate connection and resource pool management (HTTP clients, file handles) - Evaluate I/O patterns (lazy vs eager loading, streaming vs buffering, payload size, compression) - Check for unnecessary network round-trips and missing connection reuse 3. **Review Concurrency Resource Efficiency and Caching Strategy** - Assess lock granularity and contention potential — are locks held longer than necessary? - Evaluate thread pool and worker pool sizing - Identify unnecessary serialisation of independent async operations - Assess caching strategy — what to cache, invalidation approach, TTL appropriateness - Check for cache stampede / thundering herd potential **Boundary note**: System-level scalability (how the architecture handles 10x load, horizontal scaling, component failure) is assessed by the architecture lens. Resilience patterns (retry strategies, circuit breakers, timeout policies) are also assessed by the architecture lens — this lens focuses on whether the *implementation* of these patterns is efficient, not whether the strategy itself is appropriate. This lens focuses on *code-level performance* — whether individual components, algorithms, and data paths are efficient. Observability infrastructure (structured logging, metrics collection, tracing) is assessed by the code quality lens. This lens may note *what to measure* for performance but does not assess the observability design itself. Database query performance, N+1 patterns, index fitness, and migration locking are assessed by the database lens. This lens retains algorithmic efficiency and general resource management. Concurrency *correctness* (race conditions, deadlocks, data races) is assessed by the correctness lens. This lens retains concurrency *resource efficiency* (lock contention, thread pool sizing). ## Key Evaluation Questions **Algorithmic efficiency** (always applicable): - **Algorithmic complexity**: What is the time/space complexity? Is it appropriate for the expected data sizes? Are there O(n²) patterns that could be O(n) or O(n log n)? - **Data structure fitness**: What access pattern does this data structure optimise for, and does that match how it's actually used? - **Hot path efficiency**: If this code path runs 1000 times per second, which operations inside it would dominate the cost? (Watch for: allocations in loops, redundant lookups, computations that could be hoisted or cached.) **Resource and I/O efficiency** (when the change opens connections, makes network calls, or handles file I/O): - **Resource management**: What happens to this resource if the operation fails halfway through — will it be released? (Watch for: missing cleanup in error paths, connection leaks, unbounded pool growth.) - **I/O efficiency**: If the response payload grew 100x, would this code still work efficiently? (Watch for: missing pagination, unbatched network calls, eager loading of large data sets.) **Concurrency efficiency** (when the change uses threads, async/await, or shared mutable state): - **Concurrency efficiency**: What is the contention cost of the synchronisation strategy? (Watch for: coarse-grained locks that serialise independent operations, oversized or undersized thread pools, unnecessary serialisation of async work.) **Caching** (when the change involves repeated lookups or high-frequency access patterns): - **Caching**: Which operations are repeated with the same inputs, and what would the cost/benefit of caching them be? (Watch for: missing cache invalidation, inappropriate TTLs, cache stampede potential.) ## Important Guidelines - **Explore the codebase** for existing performance patterns and conventions - **Be pragmatic** — focus on performance issues that will matter at expected scale, not micro-optimisations - **Rate confidence** on each finding — distinguish measured bottlenecks from potential concerns - **Consider the data scale** — an O(n²) loop over 5 items is fine; over 50,000 items it's critical - **Check for existing optimisations** — understand whether seemingly inefficient code has already been profiled and is adequate - **Assess proportionally** — a background job doesn't need the same optimisation scrutiny as a hot API endpoint - **Suggest measurement** — when uncertain, recommend profiling rather than speculative optimisation ## What NOT to Do - Don't review architecture, security, code quality, standards, test coverage, usability, documentation, database, correctness, compatibility, portability, or safety — those are other lenses - Don't assess database query correctness, schema design, or migration safety — that is the database lens. This lens retains algorithmic efficiency and general resource management - Don't assess concurrency correctness (race conditions, deadlocks, data races) — that is the correctness lens. This lens retains concurrency *resource efficiency* (lock contention, thread pool sizing) - Don't recommend premature optimisation — only flag issues proportional to expected scale and frequency - Don't micro-optimise — focus on algorithmic and structural improvements, not shaving nanoseconds - Don't assess system-level scalability (horizontal scaling, load balancing, component failure) — that is the architecture lens - Don't assess observability infrastructure (logging, metrics, tracing design) — that is the code quality lens - Don't penalise code that has been profiled and shown to be adequate Remember: You're evaluating whether code will perform well under expected load — efficient algorithms, appropriate resource management, efficient concurrency, and effective caching. The best performance work targets the right bottleneck with the simplest fix.