Baton

Why it is shaped this way

Relay's front end, pinned, behind our driver

Status: accepted, 2026-10-02. Serves The compiler decides and Relay's words. Reopen when the pinned revision cannot be advanced for a feature we need, or stops building on stable Rust; the remedies, in order, are vendoring the crates and writing our own front end behind the same plan IR. Apollo's compiler is not a remedy.

Status: accepted, 2026-10-02. Serves The compiler decides and Relay's words. Reopen when the pinned revision cannot be advanced for a feature we need, or stops building on stable Rust; the remedies, in order, are vendoring the crates and writing our own front end behind the same plan IR. Apollo's compiler is not a remedy.

Context#

The compiler needs a GraphQL parser, a schema, a typed IR, spec validation and Relay's transforms (fragment arguments, connections, refetchable queries, identity fields, flattening). Relay's compiler is a Rust workspace, MIT, not published to a registry, versioned 0.0.0, with its output languages and source extraction closed to extension.

Decision#

Depend on Relay v21.0.1's graphql-syntax, schema, relay-schema, graphql-ir, relay-config, relay-transforms, graphql-text-printer, common and intern as git dependencies pinned to the tag. Everything in front of them (finding GraphQL in Swift and Kotlin sources) and behind them (the plan IR, the emitters, the artifact format) is ours. No fork.

Evidence#

  • Spike S1 (2026-10-02): the pipeline builds on stable Rust; the stripped binary is 4.3 MB; 500 fragments compile in 8 ms warm; Relay's own error for a misspelled field lands on the exact character inside a Swift literal.
  • Relay.swift depended on the JavaScript compiler's plugin API and was stranded when Relay 13 removed it; rescript-relay forks the Rust compiler and was 675 commits behind upstream in October 2026.
  • Baton runs its own runtime, so a pinned front end cannot fall out of step with anything; upgrades are elective.
  • The owner's web products run Relay; one directive vocabulary and one compiler lineage keep web and native in one language.

Not chosen#

  • apollo-compiler plus our own transforms: published and versioned, but a second vocabulary and a second lineage.
  • A fork: upstream tracking for no benefit, since the runtime is ours.
  • An own front end from the start: months of validation work before the first lens, measured by Isograph's two and a half years to parity.