1 oracle.pipelines.generation_progress

oracle.pipelines.generation_progress

Project validated generation attempts into durable JSON and terminal progress.

This is disposable operational evidence, never admission authority. ETA uses complete parent attempt cycles (including exclusions and setup), and reports an empirical duration spread rather than a calibrated prediction interval.

1.1 Classes

Name Description
QueueProgressReporter Keep JSON alive during a blocking worker and optionally render its counts.

1.1.1 QueueProgressReporter

oracle.pipelines.generation_progress.QueueProgressReporter(
    root,
    identity,
    *,
    planned,
    population_size=None,
    render=False,
)

Keep JSON alive during a blocking worker and optionally render its counts.

The scheduler owns the records supplied to update. A heartbeat never scans stores or claims completion; it only advances elapsed wall time for the same active assignment. Only the scheduler writes terminal outcomes.

1.1.1.1 Methods

Name Description
update Persist one scheduler transition and render exactly those counts.
close Stop heartbeat before releasing the scheduler lock.
1.1.1.1.1 update
oracle.pipelines.generation_progress.QueueProgressReporter.update(
    records,
    status,
    **fields,
)

Persist one scheduler transition and render exactly those counts.

1.1.1.1.2 close
oracle.pipelines.generation_progress.QueueProgressReporter.close()

Stop heartbeat before releasing the scheduler lock.

1.2 Functions

Name Description
estimate_remaining Estimate finite backlog from complete cycles with explicit support.
project_progress Return the single count/time projection consumed by JSON and tqdm.
read_queue_progress Read a freshness-aware operational projection without rescanning Zarr.

1.2.1 estimate_remaining

oracle.pipelines.generation_progress.estimate_remaining(
    records,
    *,
    remaining,
    active_elapsed=0,
    active=False,
)

Estimate finite backlog from complete cycles with explicit support.

Records must come from an admitted recipe queue whose frozen roster has exactly one root key per physical scene. Queue admission and ledger replay establish this caller precondition; arbitrary snippet keys cannot stand in for independent scenes. Old receipts containing only worker duration do not establish full-cycle cost. Forward errors use earlier attempts from other scenes only, leaving the predicted scene out of its training sample. These errors are descriptive backtests; no prediction-interval coverage is claimed.

1.2.2 project_progress

oracle.pipelines.generation_progress.project_progress(
    identity,
    records,
    status,
    planned=None,
    active_unit=None,
    phase=None,
    active_started_at=None,
    stop_requested=False,
    population_size=None,
    now=None,
)

Return the single count/time projection consumed by JSON and tqdm.

1.2.3 read_queue_progress

oracle.pipelines.generation_progress.read_queue_progress(
    queue_path,
    *,
    stale_after_seconds=120,
)

Read a freshness-aware operational projection without rescanning Zarr.

The writer appends ledger outcomes only after full child validation. This polling seam revalidates frozen queue identity, ledger structure, lifecycle counts and read currentness; it does not rehash completed payloads. Use read_validated_queue_snapshot at integrity/admission boundaries. Missing/stale heartbeats suppress active claims and ETA.

Back to top