Civis

Fixed
- Correct prediction template id for CivisML 1.0 (278)
- `civis.ml.ModelFuture.table` checks for primary key before reading in
data. (276)

Added
- Test for Python 3.7 compatibility (277)

Changed
- Updated mock API specs (281)

Fixed
- `civis.io.civis_to_file` will now retry on S3 connection errors (273)
- Buffers will be reset appropriately on connection failures during
`civis.io.file_to_civis` (273)

Fixed
- `_stash_dataframe_as_csv` in `civis/ml/_model.py` now uses a `StringIO`
object which has the `getvalue` method (required by `pandas` v0.23.1
if a file-like object is passed into `df.to_csv`). (259)
- `civis_to_multifile_csv` fully respects the `client` keyword argument

Added
- Added instructions in the README for adding an API key to a Windows 10
environment
- Configured Windows CI using AppVeyor. (258)

Changed
- Coalesced `README.rst` and `index.rst`. (254)
- joblib documentation has moved to readthedocs. (267)

Fixed
- Added more robust parsing for tablename parsing in io. You may now
pass in tables like schema."tablename.with.periods".
- Adding in missing documentation for civis_file_to_table
- Include JSON files with pip distributions (244)
- Added flush to `civis_to_file` when passed a user-created buffer,
ensuring the buffer contains the entire file when subsequently read.
- Fix several tests in the `test_io` module (248)
- Travis tests for Python 3.4 are now restricted to pandas<=0.20, the
last version which supported Python 3.4 (249)

Added
- Added a utility function which can robustly split a Redshift schema name
and table name which are presented as a single string joined by a "." (225)
- Added docstrings for `civis.find` and `civis.find_one`. (224)
- Executors in ``futures`` (and the joblib backend, which uses them) will now
add "CIVIS_PARENT_JOB_ID" and "CIVIS_PARENT_RUN_ID" environment variables
to the child jobs they create (236)
- Update default CivisML version to v2.2. This includes a new function
``ModelPipeline.register_pretrained_model`` which allows users to train
a model outside of Civis Platform and use CivisML to score it at scale (242, 247).
- Added a new parameter ``dvs_to_predict`` to ``civis.ml.ModelPipeline.predict``.
This allows users to select a subset of a model's outputs for scoring (241).
- Added `civis.io.export_to_civis_file` to store results of a SQL query
to a Civis file
- Surfaced `civis.find` and `civis.find_one` in the Sphinx docs. (250)

Changed
- Moved "Optional Dependencies" doc section to top of ML docs, and
added clarifications for pre-defined models with non-sklearn
estimators (238)
- Switched to pip install-ing dependencies for building the documentation (230)
- Added a merge rule for the changelog to .gitattributes (229)
- Default to "all" API resources rather than "base".
- Updated documentation on algorithm hyperparameters to reflect changes with
CivisML v2.2 release (240)

Added
- Added a script for integration tests (smoke tests).

Fixed
- Added missing string formatting to a log emit in file multipart upload and
correct ordering of parameters in another log emit (217)

Changed
- Updated documentation with new information about predefined stacking
estimators (221)
- Updated CivisML 2.0 notebook (214)
- Reworded output of `civis notebooks new` CLI command (215)

Added
- Documentation updated to reflect CivisML 2.1 features (209)
- ``civis.io.dataframe_to_civis``, ``civis.io.csv_to_civis``, and ``civis.io.civis_file_to_table`` functions now support the `diststyle` parameter.
- New notebook-related CLI commands: "new", "up", "down", and "open".
- Additional documentation for using the Civis joblib backend (199)
- Documented additional soft dependencies for CivisML (203)

Changed
- Changed `ModelPipeline.train` default for `n_jobs` from 4 to `None`,
so that `n_jobs` will be dynamically calculated by default (203)
- Use "feather"-formatted files to send data from users to CivisML, if possible.
Require this when using ``pd.Categorical`` types, since CSVs require us to
re-infer column types, and this can fail. Using feather should also give a
speed improvement; it reads and writes faster than CSVs and produces smaller files (200).
- ``ModelFuture`` objects will emit any warnings which occurred during their
corresponding CivisML job (204)
- Removed line setting "n_jobs" from an example of CivisML prediction.
Recommended use is to let CivisML determine the number of jobs itself (211).
- Update maximum CivisML version to v2.1; adjust fallback logic such that users get
the most recent available release (212).

Fixed
- Restored the pre-v1.7.0 default behavior of the ``joblib`` backend by setting the ``remote_backend``
parameter default to 'sequential' as opposed to 'civis'. The default of 'civis' would launch additional
containers in nested calls to ``joblib.Parallel``. (205)
- If validation metadata are missing, ``ModelFuture`` objects will return ``None``
for metrics or validation metadata, rather than issuing an exception (208)
- Allowed callers to pass `index` and `encoding` arguments to the `to_csv` method through `dataframe_to_civis`.

Performance Enhancements
- ``civis.io.file_to_civis`` now uses additional file handles for multipart upload instead of writing to disk to reduce disk usage
- ``civis.io.dataframe_to_civis`` writes dataframes to disk instead of using an in memory buffer