Skip to content

MIGRATING-0.2-TO-0.3.md

Latest commit

 

History

History
45 lines (33 loc) · 1.94 KB

MIGRATING-0.2-TO-0.3.md

File metadata and controls

45 lines (33 loc) · 1.94 KB
Raw

Migrating from Parzen 0.2 to 0.3

Parzen 0.3 preserves the public Rust API and tested TPE behavior of 0.2, so no source-level API migration is required. Two build and reproducibility changes may require attention.

Rust version

The minimum supported Rust version is now 1.89. Upgrade the Rust toolchain before resolving Parzen 0.3.

Default SIMD feature

The new simd feature is optional and enabled by default alongside serde. It uses Pulp for runtime-dispatched SIMD without exposing SIMD types in Parzen's public API. Supported machines select an appropriate implementation at runtime, and all other targets use the scalar fallback.

To retain Serde while using the scalar implementation from Parzen 0.2, configure the dependency as:

parzen = {
    version = "0.3",
    default-features = false,
    features = ["serde"],
}

Use default-features = false without an explicit feature list when neither Serde nor SIMD is needed.

Parzen applies the SIMD acquisition path only to at least four independent continuous estimators, the envelope where it passed the project's runtime and non-regression gates. One-dimensional, grouped, categorical, discrete, and mixed estimators retain exact scalar paths. SIMD is an internal runtime optimization and does not change the configured kernels, candidate count, history policy, or optimization budget.

Seeded reproducibility

Parzen remains deterministic for a fixed optimized build, seed, target architecture, search space, and observation order. Semantically equivalent batching and floating-point evaluation order may change the exact suggestion sequence between Parzen minor versions or target architectures. Do not use a cross-version suggestion sequence as a persistent wire or snapshot format.

The complete fixed 32-seed quality suite used for the 0.3 performance work found no accepted optimization-quality regression. This is a statistical quality result, not a promise of identical candidate sequences.