Smif

=============

Changes to tested Python and Node.js versions:

- Bump Python versions (drop 3.5 from testing, test on 3.6 and 3.8)
- Bump Node version (test on current LTS release, 14.15.1)

Fixes:

- Update npm dependencies

=============

Fixes:

- Move unit registration to within UnitAdaptor.simulate (fixes running with custom units)
- Update npm dependencies

===========

Features:

- Add smif CLI subcommands: decide, before_step, step
- Add --dry-run option when running a whole model run - prints a list of step subcommands
- Allow concrete instances of Model, SectorModel
- Include only sector_models and model_dependencies in job graphs (reduces logging noise)
- Include hundredths of a second in log timings

Fixes:
- Handle one-dimensional DataArray to Dataframe conversions (could mis-sort data with unordered
labels in conversion to/from xarray - see pydata/xarray3798)
- Update npm dependencies
- Relax ruamel.yaml requirement to >=15.50
- Fix Travis/Appveyor tests under Python 3.5 (install/pin dependencies)

=============

Fixes:

- Fix warm start re-run
- Update npm module dependencies in React app
- Fix docs build

===========

Features:

- Warm start avoids re-running models within a partially-complete timestep
- DAFNI scheduler plugin
- Convert CSV to Parquet store
- Create model runs corresponding to all variants in a scenario
- Results object gives access to all model run results (helper wrapper around a Store)

Fixes:

- Handle missing dims in dependency validation
- Clear results before new run
- Update pytest and fix testing for error messages
- Fix spec/data mismatch when reading scenario data

Config update may be required:

- Data file references in YAML configs should not include extensions (`.csv` or `.parquet`) in
order to support conversion between data store implementations.

===========

Features:

- DecisionModule now handles the state of the system of systems
- correctly; refactored pre-spec planning out of DecisionModule
- Results_Handle provides read-only access to state
- Can read and write empty state to and from store without errors
- Added documentation on how to use pre-specified planning
- Find out model runs which have results available
- Identify missing results
- Programmatically access results