Forcing and suppressing re-runs

Task recipes for taking control of the workflow when the default resume behavior is not what you want. For the why behind the default behavior, see When and why re-runs happen.

Force a full re-run of specific scenarios

Pass override_force_rerun to analysis.run(). This performs a two-layer invalidation — it deletes the _status/*.flag markers and clears the per-scenario processing-log records — so the targeted steps genuinely re-execute rather than being skipped by a surviving completion record:

# Re-run everything from scratch of the completion state:
analysis.run(override_force_rerun="all")

# Re-run only specific scenarios (by event index):
analysis.run(override_force_rerun={"event_iloc": [0, 2]})

# For a sensitivity analysis, target specific sub-analyses:
analysis.run(override_force_rerun={"sa_id": ["sa_3", "sa_7"]})

override_force_rerun accepts "all", "none", or a dict keyed by "event_iloc" / "sa_id". It overrides the config's force_rerun field for a single invocation without mutating the config.

Add scenarios or events to a running sweep

Re-run with from_scratch=False (the default). Under the toolkit's graceful-rerun semantics, analysis.run() picks up the newly added scenarios and resume-sweeps only the additions — completed scenarios are not re-run, and any still-queued jobs from a previous submission are waited on rather than re-submitted:

analysis.run(from_scratch=False)

Do not reach for from_scratch=True to add work — that wipes the analysis directory and rebuilds everything from the beginning.

See re-processed results without re-simulating

When you have completed simulations and only want to regenerate the processed outputs, consolidation, or report — for example after fixing a renderer or to inspect partial results — use analysis.reprocess(). The start_with argument controls how far back the re-processing begins:

# Re-consolidate + re-render the report from existing per-scenario outputs:
analysis.reprocess(start_with="consolidate")

# Re-render the report only (fastest; re-uses the consolidated datatree):
analysis.reprocess(start_with="render")

# Re-run per-scenario processing too (raw outputs -> zarr/nc):
analysis.reprocess(start_with="process")

reprocess() never re-runs simulations — it operates on the raw outputs already on disk, so it is safe to run while queued or running simulation workers exist.