v0.10.0
ModelManager v0.10.0
Merged pull requests:
- A refused sbatch submission is not a failed simulation (#51) (@drbergman)
- Drop eight dependencies ModelManager never loads (#55) (@drbergman)
- Restorable means JLD2 can name it; stored QoI values read back as written (#56) (@drbergman)
- Compact the maintainer records to what is true now (#57) (@drbergman)
- Give the variations databases the busy timeout, and clean up a deleted run's batch tags (#60) (@drbergman)
- Fix two discrete-parameter defects: resume, and co-variation joint consistency (#63) (@drbergman)
- An in-flight generation folder is not a completed generation (#64) (@drbergman)
- Publish the paths and defaults a backend was rebuilding by hand (#65) (@drbergman)
- Do not snap a discrete coordinate to the CDF grid (#66) (@drbergman)
- deleteCalibration(delete_subs=true) deletes only the monads that run alone used (#67) (@drbergman)
- Recover a campaign whose driver process died (#68) (@drbergman)
- Make the assertions that could not fail actually able to fail (#69) (@drbergman)
- One QoI contract across the sink, sensitivity analysis and calibration (#70) (@drbergman)
- Fix the remaining calibration review findings (#71) (@drbergman)
- Draw from an ABC-SMC posterior with samplePosterior (#72) (@drbergman)
- Choose which parameters a calibration or GSA plot draws (#73) (@drbergman)
- Draw the Sobol' total-order bars behind the first-order ones (#74) (@drbergman)
- Stop the workers from starting queued specs after run fails fast (#75) (@drbergman)
- Give the default reducer the two-argument form data selects (#76) (@drbergman)
Closed issues:
- Design brief: one QoI contract across the sink, sensitivity analysis and calibration (#52)
- Design brief: recover a campaign after the driver dies; make HPC detection overridable (#53)
- Design brief: the MM/PCMM boundary — interface contract, schema ownership, XML seam, exports (#54)
- deleteCalibration(delete_subs=true) deletes monads other runs still use, unlike deleteSampling (#58)
- Test gaps: assertions whose two arms coincide, and no end-to-end run() on the SLURM path (#59)
- CDF-grid snapping distorts a discrete parameter's prior and can leave levels unsampled (#61)
- Calibration review: the remaining 19 findings, triaged (#62)