Varlink – IPC to replace D-Bus gradually in systemd

Original Article ↗ Hacker News Discussion ↗

Serialization format (JSON) choice

  • Many criticize JSON for IPC: text-based, larger payloads, expensive to parse on low-end or high-volume systems, poor numeric model (no precise universal 64‑bit ints), weak typing, and awkward handling of binary data (base64 overhead).
  • Others argue JSON parsing is extremely fast in practice, messages are small and infrequent for system IPC, and marshalling cost is negligible compared to context switches.
  • Ubiquity is a major pro: every language has JSON libraries, often in the standard library, making bindings trivial.
  • 64‑bit integer handling is contentious: some emphasize JSON/I‑JSON float constraints and real parser limits; others say large ints can be encoded as strings or that most systemd use-cases don’t actually need >2⁵³ precision.

Performance and IPC architecture

  • Many see D-Bus’s main performance problem not in serialization but in the broker: extra context switches, enforced rate limits, and multi-step “summary then details” patterns.
  • Varlink uses direct Unix sockets, fewer roundtrips, and can “upgrade” to raw connections for bulk data, which some view as a net win.
  • Skeptics question assuming “hundreds of messages/sec” is the ceiling; system buses can see heavy traffic (desktop plugins, file managers, mass events), where JSON size overhead might matter.

Alternatives discussed

  • Proposed alternatives: CBOR (or DAG‑CBOR), MsgPack, ASN.1 (PER/OER/DER), protobufs, Cap’n Proto, existing Android Binder, Wayland-style protocols, or even raw C structs over Unix sockets.
  • Long subthread debates ASN.1’s complexity vs completeness, and protobuf/Cap’n Proto’s TLV tradeoffs, extensibility, and tooling.
  • Some feel reinventing IPC instead of adopting mature alternatives is needless NIH; others note prior attempts (kdbus, BUS1) failed to land.

Debugging and observability

  • A core justification for JSON: messages are readable in strace, fitting Unix’s “text by default” ethos and making debugging dramatically easier without specialized tools.
  • Critics respond that better tooling (e.g., strace/wireshark-style decoders for binary formats) would avoid compromising on the protocol itself.

Scope, coexistence, and ecosystem impact

  • Varlink is not a bus; it’s a point‑to‑point IPC with schemas and a resolver. Some welcome its simplicity and low barrier to writing services (e.g., quick Python scripts).
  • Others worry about “yet another IPC layer” beside D-Bus: more complexity, attack surface, and long-term churn in fundamental Linux plumbing.
  • Current reality: systemd already uses varlink alongside D-Bus; coexistence is accepted, but long‑term direction and migration strategy remain unclear.