windmill/backend/COMPILATION_OPTIMIZATION.md

Backend Compilation Optimization

Summary

Feature-gated heavy dependencies that were compiled by default but only used behind enterprise/EE feature flags. This reduces the default build from 761 crates to 609 crates (20% reduction).

Changes

Root Cargo.toml

• Removed 12 unused direct dependencies: kube, k8s-openapi, aws-sigv4, aws-sdk-config, opentelemetry-proto, systemstat, globset, libloading, bitflags, memchr, quote, pep440_rs

windmill-worker/Cargo.toml

• Made optional (only needed for EE OTEL tracing proxy): hudsucker, hyper-http-proxy, hyper-tls, hyper-util, rcgen, opentelemetry-proto, prost
• Made optional (only needed for EE features): aws-config, aws-credential-types, aws-smithy-types
• Created otel_proxy feature to group the OTEL proxy deps
• Updated private feature to include otel_proxy

windmill-common/Cargo.toml

• Made optional: aws-config, aws-credential-types, aws-smithy-types, systemstat, globset
• Added AWS deps to private, parquet, aws_auth, bedrock features
• Added systemstat to private feature
• Added globset to parquet feature

windmill-api/Cargo.toml

• Made optional: aws-sigv4, aws-sdk-config, aws-credential-types, aws-smithy-types, windmill-parser-py-imports, windmill-autoscaling
• Added AWS deps to parquet and bedrock features
• Added windmill-parser-py-imports to python and agent_worker_server features
• Added windmill-autoscaling to enterprise feature

windmill-autoscaling/Cargo.toml

• Made optional: kube, k8s-openapi (only used in EE code)
• Added to private feature

parsers/windmill-parser-py-imports/Cargo.toml

• Removed unused direct dependencies: malachite, malachite-bigint (still available transitively via rustpython-parser)

Benchmarks

Default build (no features)

Metric	Before	After
Crates compiled	761	609
Notable deps eliminated	-	aws-sdk-config (9.5s), k8s-openapi (7.3s), zstd-sys (7.7s), kube-client (1.9s)

Incremental compilation (stable-state, warm cache)

Scenario	Before	After
Touch windmill-api/src/users.rs	~5.6s	~5.4s
Touch windmill-worker/src/worker.rs	~6.7s	~6.2s
Touch windmill-common/src/worker.rs (cascade)	~8.5s	~8.5s

Incremental compilation improvement from feature-gating alone is modest because the bottleneck is the compilation of the windmill crates themselves (especially windmill-api at 90k LOC), not the dependencies.

Developer-Local Speed Tips

These settings are not committed because they are developer-local preferences that depend on toolchain availability. Combined, they yield ~16% faster incremental compilation.

mold linker (~6% improvement)

Install mold and add to .cargo/config.toml:

[target.x86_64-unknown-linux-gnu]
linker = "clang"
rustflags = ["-C", "link-arg=-fuse-ld=mold"]

Reduced debug info (~10% improvement)

Add to [profile.dev] in Cargo.toml:

split-debuginfo = "unpacked"
debug = "line-tables-only"

SQLX offline mode (~4% improvement)

If you're not modifying SQL queries:

export SQLX_OFFLINE=true

Combined effect

Scenario	Baseline	With all tips
Touch windmill-api file	5.6s	4.7s
Touch windmill-worker file	6.7s	6.0s
Touch windmill-common file (cascade)	8.5s	7.6s

What would help more (future work)

The single biggest improvement would be splitting windmill-api (90k LOC) into smaller crates. Currently, any file change in the crate triggers re-analysis of all 90k lines. However, this requires significant refactoring due to tight coupling between the triggers subsystem, jobs, users, and the axum router initialization.