콘텐츠로 이동

영어 원문

이 설계 기록은 역사적 맥락과 원문의 의미를 보존하기 위해 영어로 제공합니다.

ADR-031: Monorepo split into control plane (inferplaned) and node-local data plane (mayu)

Status

Accepted (2026-07-30)

Context

inferplane began as a single central gateway binary. A central gateway sits on the inference critical path, which contradicts the project's core argument:

  • Streaming latency — a central hop taxes every SSE chunk and directly hurts TTFT (benchmarkable, the easier claim to defend).
  • Fault isolation — a central outage must not stop every developer.

The repository was renamed inferplane/mayu → inferplane/inferplane and is restructured as a monorepo with two binaries.

Decision

  1. Two binaries, one module.
  2. cmd/inferplaned — control plane: policy/routing-rule distribution, budget-lease issuance, short-lived credential brokering, telemetry aggregation. Inference traffic never passes through it.
  3. cmd/mayu — node-local data plane (localhost or K8s node-local), the existing gateway code moved via git mv. mayu is a component name, not a project name — the same position ztunnel/waypoint hold in Istio.

  4. cmd/inferplaned is scaffolded now (health endpoints only) so both binaries import internal/policy from day one and schema version skew is a compile error, not a production surprise.

  5. Hybrid config API (api/v1alpha1). The schema borrows the CRD shape (apiVersion/kind/metadata/spec — kubectl-friendly, CNCF-idiomatic) but delivery is inferplane's own gRPC/HTTP channel: workstation mode has no K8s API server, so the types depend on no Kubernetes machinery.

  6. Budget control is a lease pattern. N proxies each see only their own usage, so rule propagation alone cannot enforce a global budget. The control plane issues leases ("this much budget for this interval"); the data plane enforces locally with zero network round trips inside the grant, reports consumption and renews asynchronously. ★ Open: lease issuance unit and renewal cadence defaults — the single parameter pair trading overshoot tolerance against control-plane load. Config fields exist and are required; no default is hardcoded.

  7. Failure policy is per rule, never global (failurePolicy on every rule, required, no default). Control-plane outage → inference continues within lease validity (fail-open); only hard budget caps fail closed on lease expiry. Global fail-open voids budget control; global fail-closed voids the fault-isolation argument.

  8. Version skew is explicit. A data plane receiving a rule it does not support rejects it (policy.UnsupportedError) and reports the rejection; the control plane exposes the version distribution of connected proxies. Silent ignoring is the worst failure mode a governance tool can have.

  9. Local cache never replaces the server-side prompt cache. Its three jobs: identical-request dedupe, cache-affinity routing (pin a session to the same region/inference profile to keep the server cache warm — the most important), offline spooling. Payloads live only in a VolatileStore (memory; named for the guarantee, not tmpfs — macOS has neither tmpfs nor /dev/shm); lease state / counters / prefix hashes / routing decisions live in SQLite and never contain payloads.

  10. Cache affinity vs fallback is an explicit per-rule preference (onAffinityConflict): failover cold-starts the server cache and can raise cost, so neither side wins by default.

  11. Credentials never rest in the data plane. The control plane authenticates users via OIDC and brokers short-lived credentials (e.g. STS AssumeRole), which also anchors per-user attribution at issuance time. Standalone (control-plane-less) mode may use local config keys but must print a "not recommended for production" warning at runtime.

Consequences

  • Provider isolation (§8), the canonical-schema invariant (§2.2), and the verbatim-body cache invariant (§4.4) are unchanged and now live in mayu.
  • Existing internal/router/server/openai/tracing/metrics keep serving the static-config path; internal/proxy/cache/telemetry are the consolidation targets for the lease-driven path.
  • Old release links (v0.1.0/v0.2.0) pointed at tags that were never pushed; versioning restarts when the first split release is cut.

Open questions (★ — decide before implementing)

  1. ~~Lease issuance unit and renewal cadence defaults~~ — decided in ADR-032 (grant 0.1% of limit, renew 10s; policy poll 60s, floor 15s).
  2. ~~api/v1alpha1 shape~~ — decided: hybrid (CRD-style schema, gRPC/HTTP delivery).
  3. Whether to track server-cache TTL with a proxy-local timer (providers do not return remaining TTL; knowing expiry opens a window to switch to a cheaper path, but complexity vs. benefit is unjudged).