v0.4.2

Highlights

Faster Kubernetes log compression

The kubernetes-pattern detector was recompiling its 14 regex patterns from
scratch on every input line — a hot allocator burn that turned what should
be microsecond-per-line work into the dominant cost on K8s-heavy logs.
Patterns are now cached at startup, the same way every other detector in
the codebase already worked. Kubelet, argocd, rancher, cilium, and similar
logs compress 4-6× faster on dense inputs; logs with no Kubernetes content
are unchanged.

Faster Linux release binary

The released Linux x86_64 binary is a single-file static musl build that
runs anywhere — Alpine, distroless, RHEL 7. musl's default malloc is
substantially slower than glibc's under multi-threaded allocation
pressure; v0.4.2 statically links mimalloc as the global allocator on
musl, recovering glibc-class throughput while keeping the single-file
standalone distribution. Combined with the Kubernetes-pattern fix above,
parallel compression of kubelet logs is 79× faster end-to-end vs v0.4.1.

---

0.4.2 (2026-05-03)

Performance

• cache kubernetes detector regexes in LazyLock (9779568)
• use mimalloc as global allocator on musl (88e4f04)

Full changelog: v0.4.1...v0.4.2