Skip to content

Discovery and execution

The lifecycle begins with a stable project/source identity and ends only when control state, external observations, artifacts, and provenance explain the outcome.

End-to-end flow

Discovery

Discovery is a repeatable metadata preparation pipeline, not an execution side effect.

Step Project control Durable result
Build query discovery.queries and source template variables Rendered archive request context
Enrich discovery.enrichments Additional archive rows
Prepare fields prepare_metadata.field_map and typed transforms Normalized metadata
Derive flags prepare_metadata.discovery_flags Admission-relevant facts
Sign prepare_metadata.signature Stable change digest

Exclude volatile archive fields such as access URLs, sizes, and timestamps from the signature when they should not trigger new scientific work.

Admission and preparation

Admission checks automation policy, source readiness, queue pressure, and per-profile concurrency before creating runnable work. Preparation then pins the project revision and deployment-profile snapshot, renders the manifest, applies graph patches, and records checksums.

configrevision profilesnapshot manifestsha256 graphsha256

Submission and reconciliation

Submission intent is durable before I/O. Slurm job identity and DALiuGE session identity are persisted as soon as they are known. Pollers then update scheduler and DALiuGE axes independently; the reducer derives the next safe control action.

Observation Meaning
Scheduler running, DALiuGE building Normal startup progression
Scheduler succeeded, DALiuGE running Inconsistent; investigate rather than complete
SSH disconnected after sbatch Submission uncertain; search before retry
DALiuGE finished, outputs unverified Continue output verification
Outputs verified Eligible for successful terminal outcome

Operator reading order

For a stalled execution, inspect:

  1. Readiness and admission diagnostics.
  2. Execution control phase and external axes.
  3. Worker claim and heartbeat.
  4. Scheduler job and DALiuGE session identifiers.
  5. Immutable artifacts and run record.
  6. Provenance timeline and metrics around the transition.

The execution state model defines exact values and retry gates. Deployment profiles define the backend behavior pinned to each execution.