0006. Pipeline ontology: flat recursive DAG

• Status: Accepted
• Date: 2026-07-12
• Deciders: thulasi.ram (architect)

Context

Incumbents fix a tier count: GHA Workflow→Job→Step, Woodpecker Pipeline→Workflow→Step, Concourse Pipeline→Job→Step. That middle tier exists because “job = pod, steps = sequential-in-pod sharing a filesystem.” We dissolved that rationale in 0004: Step is the Pod boundary, and the workspace passes via content-addressing. So parallelism no longer needs a “job” tier (it is emergent from the DAG), and reuse doesn’t need two mechanisms.

Decision

A flat recursive DAG:

• Pipeline → Step, with dependencies expressed directly between steps (needs:). Parallelism is emergent (non-dependent steps run concurrently, each its own Pod).
• Composition = recursion: a Step can invoke another Pipeline. This single concept replaces “reusable workflows” and “composite actions” and Woodpecker’s “workflow” tier. A “job” is just a named subgraph — sugar, not structure.
• Matrix is an orthogonal modifier on any Step (fan-out), not a tier (0023).

Consequences

• Fewer concepts, strictly more expressive power; maps 1:1 onto the durable DAG state machine and the pod-per-step executor.
• Unfamiliar to GHA migrants (no jobs: key) — mitigated later by a GHA-import shim.
• invoke (same-run nesting) is distinct from on: upstream (new run); see 0025.

Alternatives considered

• Two-tier Pipeline→Job(DAG of Steps) — GHA-familiar, but the tier is vestigial once step = pod, and reuse still needs a second mechanism.
• Three-tier Pipeline→Workflow→Step — most redundant given our model.