The Pipeline Framework

Appearance

Protobuf-over-HTTP Dispatch Design (Orchestrator) ​

Scope ​

This document defines orchestrator-side dispatch changes required to make Protobuf-over-HTTP replay-safe under at-least-once delivery and multi-instance deployments.

Status in current GA track:

  1. Contract parity is being aligned across REST, gRPC, LOCAL, and Protobuf-over-HTTP transports.
  2. Async control-plane semantics are shared across transports and platform paths (same transition identity and retry model).
  3. Event-sourced dispatch journaling remains a future evolution path, not a GA blocker.

Dispatch Metadata Generation ​

For every outbound dispatch, orchestrator must emit:

  • x-tpf-correlation-id
  • x-tpf-execution-id
  • x-tpf-idempotency-key
  • x-tpf-retry-attempt
  • x-tpf-deadline-epoch-ms (optional; when absent, receiver executes without deadline guard)
  • x-tpf-dispatch-ts-epoch-ms
  • x-tpf-parent-item-id (optional)

Idempotency keys must be deterministic from stable work identity (pipeline + step + business key + lineage index), never random UUIDs.

Durable Dispatch State ​

Current baseline in GA code:

  1. execution-level durable state is tracked in shared store (ExecutionStateStore),
  2. work dispatch is durable through queue dispatcher providers (WorkDispatcher),
  3. per-dispatch item journaling is not yet the runtime engine of record.

Future hardening target:

  • Track every outbound item in durable shared storage keyed by (executionId, stepId, idempotencyKey).

Required states:

  • PENDING
  • SENT
  • ACKED
  • FAILED_RETRYABLE
  • FAILED_FINAL

Transitions:

  • PENDING -> SENT when wire dispatch starts.
  • SENT -> ACKED on successful response commit.
  • SENT -> FAILED_RETRYABLE for retryable envelope codes.
  • SENT -> FAILED_FINAL for non-retryable envelope codes.
  • FAILED_RETRYABLE -> SENT on redelivery attempt (retry attempt incremented).

Crash Recovery ​

On restart:

  1. Scan records not in ACKED/FAILED_FINAL.
  2. Re-enqueue deterministically honoring ordering constraints.
  3. Preserve idempotency key and correlation id.
  4. Increment retry attempt only for true redelivery.

Multi-instance Coordination ​

  • Use shared durable storage (database or consistent KV) for dispatch state.
  • Leader election is optional if state transitions are CAS/transactional.
  • Duplicate in-flight sends are tolerated because operator side is dedupe-safe using stable idempotency keys.

Planned follow-up implementation work ​

  • Introduce a dispatch store SPI for pluggable per-item durability.
  • Implement an orchestrator worker reconciliation loop for orphaned SENT records.
  • Provide durable dead-letter routing for FAILED_FINAL with operator-facing diagnostics.

Additional work: dispatch parity across all transports ​

The dispatch metadata/state-machine contract must be enforced consistently beyond Protobuf-over-HTTP.

Target transports ​

  • gRPC transport
  • REST/JSON transport
  • LOCAL transport (for parity testing and deterministic simulation)

The FUNCTION platform path should preserve the same metadata contract through generated provider handlers, but it is a platform mode rather than a transport.

Required parity outcomes ​

  • Identical canonical dispatch metadata keys and semantics.
  • Preserve retry-attempt increment rules.
  • Maintain deadline evaluation behavior (absolute deadline only).
  • Preserve duplicate suppression contract (executionId + stepId + idempotencyKey).
  • Keep terminal classification into retryable/non-retryable outcomes unchanged.

Parity test slice ​

  • Add transport-parameterized tests asserting identical behavior for:
    • duplicate dispatch
    • replay after crash simulation
    • retry exhaustion handling
    • deadline-expired fast-fail