Performance Best Practices

Optimize the measured hot path only. Rust gives zero-cost abstractions, but discipline on allocations, profiling, and release settings still wins in production.

How to Use This List

• Apply before major perf work or after latency SEVs.
• Pair with Benchmarking & Profiling for tooling.
• Require benchmark or profile evidence in perf PR descriptions.

A - Measure

• Benchmark and profile release binaries. Never tune debug builds for production SLOs.
• Establish baseline metrics before changes. Criterion or APM p50/p99 documented in PR.
• Use representative workloads. Real query shapes, payload sizes, concurrency levels.
• Separate CPU vs I/O bottlenecks. Profiles empty when waiting on network/DB.
• Re-measure after each change. One knob at a time.

B - Hot Path Code

• Borrow before clone in inner loops. &str, &[T], Cow at boundaries.
• Preallocate Vec/String with with_capacity. Known sizes from input metadata.
• Prefer iterators until profile says loop. Zero-cost abstractions are real in release.
• Avoid Box<dyn> and dynamic dispatch in hot paths. Use enums or generics.
• Batch I/O and DB round trips. N+1 queries dominate many Rust APIs.

C - Async Services

• Never block Tokio worker threads. spawn_blocking for CPU/sync IO.
• Limit concurrent tasks with semaphores. Protect DB and upstream APIs.
• Use tracing spans on handler boundaries. Find slow awaits in production.
• Connection pool sized from load tests. sqlx pool not unlimited.
• Timeout all external calls. Prevent pile-up under slowness.

D - Build and Deploy

• Release profile: opt-level = 3, thin LTO, codegen-units = 1 for servers. Tune with data.
• Strip symbols in shipped artifacts; keep debug info separately if needed. Crash symbolication.
• Disable unused crate features workspace-wide. Smaller binaries and faster compiles.
• Pin target-cpu to deployment baseline. Not native unless uniform hardware.
• Track binary size on CLI/WASM releases. cargo bloat in release checklist.

E - Data and Libraries

• Use Polars lazy queries; avoid premature collect(). Polars 0.46+ vectorizes when possible.
• Stream large CSV/Parquet instead of loading all. Memory bounds latency tails.
• serde borrow from input buffer when parsing. #[serde(borrow)] on &str fields.
• Choose HashMap vs BTreeMap deliberately. Cache and ordering tradeoffs.
• SIMD only after scalar baseline profiled. Complexity tax is real.

FAQs

Related

• Performance Basics - introduction
• 30 Performance Rules - quick rules
• Release Tuning - profile flags
• Avoiding Clones - ownership

Stack versions: This page was written for Rust 1.97.0 (edition 2024), Tokio 1.x, Axum 0.8, serde 1.0, sqlx 0.8, clap 4, and Polars 0.46+.