Bypassing the kernel for 56ns cross-language IPC

Quick take

It has been reported that Tachyon — a new interprocess communication effort — documents architecture decisions designed to minimize kernel involvement and, allegedly, hit a jaw‑dropping 56ns cross‑language IPC. The project's docs live as Architecture Decision Records (ADRs) on GitHub; each ADR spells out a trade-off, the reasoning and the consequences. Accepted entries dated 2026‑03‑30 show the team has already settled several core choices rather than leaving them to ad‑hoc experimentation.

What the ADRs specify

The ADR index reads like a checklist for raw speed: memfd_create vs shm_open; SPSC strict vs MPSC queue designs; futex vs eventfd for consumer sleep; a 64‑byte message alignment; SCM_RIGHTS vs named shared memory; and a "no‑serialization contract" that avoids copying and serialization across language boundaries. These are accepted decisions — not tentative proposals — and the repo follows a strict template so decisions are auditable and immutable once merged. It has been reported that these choices are expressly aimed at removing kernel round trips and forcing direct shared‑memory handoffs between language runtimes.

Why this matters

Why chase 56ns? Because in latency‑sensitive systems, every nanosecond is money, user experience, or the difference between a queueing hiccup and graceful scaling. Kernel bypass isn't new — think DPDK or RDMA in networking — but applying the same philosophy to general cross‑language IPC is a bit of an arms race for developers juggling microservices, mixed‑language stacks, and high‑frequency data pipelines. The emotional punch here: developers get closer to the metal. Fast, yes — but also delicate.

Caveats and next steps

Benchmarks and real‑world validation matter. The 56ns claim is alluring, but it has been reported that independent verification is still pending; benchmarks can hide assumptions about CPU pinning, cache effects, and privileged setup. Kernel‑bypass brings portability, security, and maintenance trade‑offs — and complexity — so adopter caution is warranted. Still, the ADRs give a clear map for others to replicate or critique the approach. Fast and furious, but handle with care.

Sources: github.com/riyaneel, Hacker News