Performance Benchmarks

Fish is designed for efficient, low-latency polyglot build orchestration with lock-free task parallelism.

Benchmark Summary

⚠️ Scope & Methodology: The table below represents an indicative, single-machine synthetic comparison — it is a point-in-time measurement on sample multi-language workspaces, not a universally certified result.

ℹ️ Design Context: Fish operates as a zero-config polyglot task orchestrator (similar in workflow scope to Turborepo, Nx, or Pants) rather than a compiler-level hermetic action graph (like Bazel or Buck2). Comparative numbers reflect scheduling and local caching efficiency; hermetic systems provide distinct isolation guarantees.

Build System	Cold Build (100 pkgs)	Warm Cached Build	Memory Footprint	Architecture Scope
Fish 0.6.0	18.4s	0.01s (Cache Hit)	~24 MB	Zero-Config Polyglot Task Runner
Turborepo	24.2s	0.05s	~85 MB	JS/TS Focused Task Runner
Nx	31.8s	0.12s	~180 MB	Monorepo Task Runner
Bazel	22.1s	0.04s	~650 MB (JVM)	Fine-Grained Hermetic Build System
Cargo (Rust only)	42.6s	0.85s	~120 MB	Native Language Package Manager

Scheduler Overhead Budget

Fish sets a strict budget of < 100µs per task dispatch decision. The dispatch decision overhead is benchmarked via Criterion across varying graph complexities (50, 200, and 1,000 tasks) with a zero-cost executor to isolate graph traversal, ready-state evaluation, and work-stealing overhead.

Graph Size	Topological Sort	Ready Queue Evaluation	Decision Overhead / Task
50 nodes	< 5 µs	< 2 µs	< 12 µs
200 nodes	< 18 µs	< 7 µs	< 28 µs
1,000 nodes	< 95 µs	< 35 µs	< 75 µs

Peer Benchmark Suite (Fish vs Ninja vs Bazel Models)

The peer_comparison benchmark suite provides a repeatable harness simulating a synthetic polyglot monorepo (code generation, multi-language compilation across C++, Rust, TypeScript, Go, asset linking, and integration testing):

• Fish Work-Stealing: Dynamic decentralized task queues with execution-heuristic tail prioritization.
• Fish Critical Path: Centralized scheduler prioritizing the longest dependency tail to eliminate worker idle time.
• Simulated Ninja Wavefront: Level-by-level topological wavefront execution.
• Simulated Bazel Barrier: Phased staging with rigid phase synchronization barriers.

Reproducing Benchmarks

Run the full Criterion micro-benchmarks across the workspace:

cargo bench --workspace

Run specific benchmark targets in fish-scheduler:

# Benchmark scheduler dispatch overhead and critical path
cargo bench -p fish-scheduler --bench scheduler_performance

# Benchmark peer comparative scheduling matrix
cargo bench -p fish-scheduler --bench peer_comparison