All notable changes to Henosis will be documented in this Changelog.
The format is based on Keep a Changelog and adheres to Semantic Versioning.
- Added a check that resets
models.preload_picklesto False if the AWS S3 bucket is empty for consistency (issue #23). - Added
load_modelandload_generatorsfunctions, which allow users to load a Models().SKModel() object from a running system and the generator functions associated with that object, respectively (issue #1). - Added
deletefunction, which allows users to delete a Models.SKModel() object from a running system and the generator functions associated with that object (if not shared with other models), respectively (issue #1).
- Migrated to Flask v1.0.2 from 0.12.2 (issue #17).
- Changed the manner in which a connection object is referenced when calling
Server()to match the behavior when modeling (issue #15). - The manner in which API and Elasticsearch auth is specified in
config.yamlto correspond with the fix related to issue #22.
- A behavior whereby an error in retrieving models from AWS S3 caused the application / API server to crash. Added a while loop to check for success and continue attempting to retrieve models from S3 until successful (issue #19).
- An issue where an AWS S3 connection was being instantiated twice on server instantiation.
- An issue that was causing a timer start even if
models.refresh_picklesis set to None or False. - A security vulnerability that was opened by using
yaml.loadas opposed toyaml.safe_load(issue #22).
- A small unintended KeyError introduced when logging requests after the 0.0.9 update.
- A check that creates a temporary directory needed when storing models or tagging functions.
- Additional fields for the response from the /requestlog API endpoint, providing a "modelsUsed" field and a "modelsWithheld" field. When predict_probabilities is set to True in your configuration file, these fields provide context behind how frequently models meet their minimum recommendation (confidence) threshold.
- Additional checks for errors in making predictions.
- The default model field return when requesting the API for recommendations. Changed from modelsQueried to modelsUsed.
- A deprecation warning related to the WSGI import from gevent.
- An issue with relative imports across Henosis files, changed to absolute imports.
- Added the ability to set a minimum confidence threshold that dictates when individual models will provide recommendations, which can be useful in certain use cases.
- Added in a timer that periodically reloads models in memory if preload_pickles is set to True and a float or integer value (in minutes) is specified in the instance's configuration file.
- Added a check to ensure that only models in the pickle_jar are checked for use if preload_pickles is set to True. This ensures no error is encountered if a model has been deployed but isn't yet in the pickle_jar.
- Included additional logging for improved system monitoring and debugging.
- Restructured Henosis such that server deployment and modeling classes are separated and can be imported into a Python script or application with different imports. This allows for easier use of Henosis by users and for better code maintainability. Breaks previous functionality.
- An issue that was causing an error when using scikit-learn's TfidfVectorizer.
- Removed pymssql requirement and functionality.
- An issue where all X and y data (not only X_train and y_train) were resampled when upsampling or downsampling, causing the test or validation set information to bleed into the training data.
- A KeyError issue that prevented request logs to be stored in certain scenarios.
- Cross-origin request configuration and cleanup of code base.
- Cross-origin request configuration.
- Addition of Flask-CORS to allow for cross-origin requests.
- Initial release of Henosis codebase on PyPi.