FAQ¶

Frequently asked questions about ArqonHPO.

General¶

What is ArqonHPO?¶

ArqonHPO is a microsecond-budget hyperparameter optimizer written in Rust with Python bindings. It's designed for systems where optimization happens as a continuous control loop, not a one-time search.

When should I use ArqonHPO vs Optuna?¶

Is it production-ready?¶

Yes. ArqonHPO is used in production for:

• LLM inference batch sizing
• Real-time control loops
• SRE automation

We have 91% test coverage and a strict Constitution.

Algorithm¶

What optimization algorithms does it use?¶

1. Nelder-Mead — Simplex method for smooth functions
2. Multi-Start Nelder-Mead — Parallel restarts for multimodal
3. TPE — Bayesian optimization for noisy landscapes

The Classify phase automatically selects the best strategy.

Is it deterministic?¶

Yes! With a fixed seed, ArqonHPO produces identical results:

# Same seed = same sequence
solver1 = ArqonSolver('{"seed": 42, ...}')
solver2 = ArqonSolver('{"seed": 42, ...}')
# solver1.ask() == solver2.ask()  ✓

How does ArqonProbe work?¶

ArqonProbe generates Low-Discrepancy Sequences for uniform parameter space coverage:

probe = ArqonProbe(config_json, seed=42)
point = probe.sample_at(0)    # First point
points = probe.sample_range(0, 100)  # Points 0-99

It's stateless and shardable — worker N can generate points [N100, (N+1)100) without coordination.

Performance¶

How fast is it?¶

Why is it faster than Optuna?¶

1. Rust core — No Python GIL, no interpreter overhead
2. Batch processing — Amortize overhead across many candidates
3. Stateless probing — No synchronization between workers

Can it run on embedded systems?¶

Yes. The CLI compiles to a ~5MB static binary with no runtime dependencies.

Safety¶

What are Guardrails?¶

Guardrails prevent dangerous configurations:

• Bounds — Absolute limits on values
• Delta limits — Max change per update
• Rate limits — Max updates per second

See Safety for details.

Does it support rollback?¶

Yes. Configure a rollback policy:

{
  "rollback_policy": {
    "max_consecutive_regressions": 3
  }
}

After 3 worse results, it reverts to the last good config.

Integration¶

How do I monitor it?¶

• TUI — arqonhpo tui --state state.json
• Dashboard — arqonhpo dashboard --state state.json
• Prometheus — --metrics-addr 127.0.0.1:9898

Can I use it with Kubernetes?¶

Yes, via:

1. CLI in a sidecar container
2. Python bindings in your app
3. Dashboard for monitoring

Helm chart planned for v0.4.

Does it work with Ray/Dask?¶

ArqonProbe is designed for distributed workers — each worker can sample independent ranges without coordination.

Troubleshooting¶

See Troubleshooting Guide¶

More Questions?¶

• GitHub Discussions
• Issue Tracker