Deployment Best Practices
Twenty habits for shipping Rust binaries and containers that are small, reproducible, and observable.
//
Search across all documentation pages
Twenty habits for shipping Rust binaries and containers that are small, reproducible, and observable.
Always cargo build --release --locked in CI: Debug artifacts must never reach production.
Commit Cargo.lock for applications: Reproducible deploys depend on a frozen dependency graph.
Pin rust-version and CI toolchain: Match MSRV across dev laptops and pipelines.
Multi-stage Dockerfiles: Compile in builder; runtime image contains binary plus CA certs only.
Run as non-root in containers: Distroless nonroot or explicit USER directive.
Prefer rustls over OpenSSL in images: Fewer native deps and smaller cross-compile surface.
Expose /health and /ready: Liveness vs readiness must differ when DB is down.
Graceful shutdown on SIGTERM: Drain in-flight requests before exit (Tokio signal handlers).
Config via environment: Twelve-factor injection; no secrets in images or repos.
Structured JSON logs: tracing-subscriber JSON for centralized search.
RED metrics on HTTP services: Rate, errors, duration histograms exported to Prometheus.
Sample distributed traces: OTLP export at 1-10% unless debugging active incident.
Cross-compile in CI matrix: Build amd64 and arm64 before users report wrong arch binaries.
Strip or LTO release binaries: Smaller images and faster cold starts.
Run migrations as separate Job: Schema changes complete before new app pods serve traffic.
Tag releases with SemVer: Container tags and git tags align (v1.2.3).
Cache CI dependencies: Swatinem/rust-cache or sccache for sub-10-minute PR feedback.
Fail CI on clippy -D warnings and fmt --check: Style and lint debt blocks merge.
Document rollback: Previous image tag retained; kubectl rollout undo tested quarterly.
Post-deploy smoke test: Curl health endpoint and one critical API path from CI or synthetic monitor.
glibc on standard Linux distros; musl for scratch/Alpine when portability beats peak allocator performance.
Rolling is default on Kubernetes; blue/green when schema migrations need instant switchback.
Aim under 50 MB compressed for single Axum service; investigate if over 100 MB without ML deps.
Keep observability agents as platform sidecars; business logic stays one Rust binary per service.
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+.
Reviewed by Chris St. John·Last updated Jul 19, 2026