ci: tooling enhancements — semver-checks, cargo-hakari, feature pruning, divan evaluation

[strawgate]

Problem

Several CI tooling and build infrastructure improvements could benefit the project:

1. cargo-semver-checks in CI

Prevents accidental breaking API changes if crate APIs ever become public.

2. cargo-hakari for workspace dependency unification

Reduces redundant compilation of shared dependencies across workspace members.

3. cargo-features-manager feature pruning

Dependencies like Arrow, DataFusion, Tokio have many feature flags. Pruning unused features can reduce compile time and binary size.

4. Evaluate divan as criterion alternative

Newer benchmarking library with simpler API and built-in allocation profiling.

5. Parallel compiler frontend

Rust 1.85+ supports parallel compilation of a single crate. Evaluate -j flag benefits.

References

• corrode.dev: Tips for Faster Rust Compile Times
• blessed.rs

[strawgate]

CI Tooling Evaluation Results (April 2026)

1. cargo-semver-checks — ✅ GO for CI

Trial results: Ran against 5 key crates (logfwd-core, logfwd-arrow, logfwd-config, logfwd-output, logfwd-types) comparing HEAD vs HEAD~20.

Crate	Checks	Pass	Skip	Breaks	Time
logfwd-core	196	196	56	0	1.4s
logfwd-arrow	196	196	56	0	19.2s
logfwd-config	196	196	56	0	13.9s
logfwd-output	196	196	56	0	33.6s
logfwd-types	196	196	56	0	6.9s

Verdict: Works out of the box with --baseline-rev. No crates.io publication needed — git baseline comparison is sufficient. Total workspace scan ~75s. Suitable for PR CI.

Setup: cargo semver-checks check-release --workspace --baseline-rev origin/main in CI.

2. cargo-hakari — ⏸️ DEFER

Requires creating a workspace-hack crate and modifying every crate's Cargo.toml to depend on it. This is an invasive change that touches every crate in the workspace. The benefit (faster builds by unifying feature resolution) is real but the setup cost is high and it adds maintenance overhead.

Verdict: Defer until build times become a bottleneck. Current sccache + default-members provides adequate caching.

3. Feature pruning — 🔍 NEEDS AUDIT

Not evaluated this session. Requires cargo-features-manager or manual audit of Arrow, DataFusion, Tokio feature flags. Arrow alone has ~30 feature flags.

Recommendation: Run cargo tree -e features -p arrow to identify unused features. Likely wins: disable Arrow IPC features not used, trim DataFusion feature set.

4. divan evaluation — ⏸️ DEFER

The project already has extensive Criterion benchmarks (scanner, pipeline, OTLP, full-chain). Migrating to divan would require rewriting all bench harnesses. divan's allocation profiling is nice but not urgent.

Verdict: Not worth the migration cost. Keep Criterion. Consider divan for new benchmark modules only.

5. Parallel compiler frontend — ✅ FREE WIN

Rust 1.94 is installed (parallel frontend stable since 1.85). Currently NOT enabled. No -Zthreads or RUSTC_PARALLEL flags set.

Fix: Add to .cargo/config.toml:

[build]
rustflags = ["-Zthreads=8"]

Or in CI: RUSTFLAGS="-Zthreads=8". Zero risk, typically 10-20% faster single-crate compilation.

Note: This flag controls parallelism within a single crate's compilation. Combined with Cargo's inter-crate parallelism, it should reduce wall-clock build time.

Summary

Tool	Verdict	Effort	Impact
cargo-semver-checks	GO	1 CI step	Prevents API breaks
Parallel compiler	GO	1 line in config.toml	10-20% faster builds
Feature pruning	Audit needed	Half day	Potentially significant
cargo-hakari	Defer	High setup cost	Moderate
divan	Defer	Migration cost	Low (Criterion works fine)

[strawgate]

Investigated all 5 items:

1. cargo-semver-checks: Skip — all 19 crates are publish = false. No public API consumers.
2. cargo-hakari: Worth doing — tokio has divergent feature sets across crates. ~10-25% cold build speedup. Filed as part of build optimization backlog.
3. Feature pruning: Moderate value — opentelemetry_sdk and tokio workspace entries lack default-features = false. Primarily compile-time, not binary size (LTO handles that).
4. divan: Skip — CI nightly benchmark pipeline is tightly coupled to Criterion JSON output (target/criterion/**/estimates.json). divan has no stable equivalent. Migration breaks the dashboard.
5. Parallel compiler: Wait — -Z threads still nightly-only. Track Fix watchOS and visionOS for pread64 and pwrite64 calls rust-lang/rust#124089 for stable stabilization.

Bonus: sccache is commented out in .cargo/config.toml — re-enabling is the single highest-impact zero-effort build speed win.

Actionable items (hakari + feature pruning + sccache) will be tracked separately.