Release Notes

v0.14.5 (2020-05-13)

New Features

• support logging Keras models in TensorFlow 2.X
• support eagerly-executed TensorFlow Tensors in ModelAPI

Bug Fixes

• filter out spaCy models when versioning pip requirements files

v0.14.4 (2020-05-04)

Backwards Incompatibilities

• run.log_training_data() no longer uploads a “train_data” artifact, and instead directly generates a histogram for deployment data monitoring

v0.14.3 (2020-04-20)

Backwards Incompatibilities

• replace commit.branch() with commit.new_branch()

New Features

• enable passing in ~ as part of filepaths
• enable setting host from $VERTA_HOST
• capture versioning information from S3

Internal Changes

• add click as a dependency in preparation for CLI development
• use back end for commit merges
• use back end for commit reverts

v0.14.2 (2020-04-01)

New Features

• use friendly default messages for merges and reverts
• implement __repr__ for Blobs

Bug Fixes

• filter out spaCy models from pip freeze
• make dataset.dataset_type friendlier
• enable e.g. Notebook Blobs to be retrieved from a Commit outside of Notebooks
• enable set_repository() without Verta authentication credentials
• validate Client config file against protobuf spec
• add more helpful typechecks on Commit methods

Internal Changes

• update protobufs for diffapi changes
• fix race condition when running parallel tests
• update tests to delete children Commits first

v0.14.1 (2020-03-17)

New Features

• add complete versioning system
• enable going directly from Client to ExperimentRun using Verta config
• add public_within_org option for set_project() and set_dataset()
• add aliases for Client’s set/get-or-create methods

Bug Fixes

• enable larger sets of ExperimentRuns to be queried
• enable ZIPing files that have invalid timestamps

Internal Changes

• move cloned dependencies to their own submodule
• move internal utils into their own submodule

v0.14.0 (2020-02-11)

Backwards Incompatibilities

• require token in DeployedModel::from_url(), even if it’s None

New Features

• add a workspace parameter to client.set_project() and client.set_dataset()
• enable logging singleton scalar NumPy arrays as metrics
• implement Keras Client integration
• implement PyTorch Client integration
• implement scikit-learn Client integration
• implement TensorFlow Client integration
• implement TensorBoard Client integration
• implement XGBoost Client intergation

Bug Fixes

• allow negative numbers in Python 3 expt_runs.find() queries

Internal Changes

• interpret HTTP 403s on getProject as project not found
• include gorilla as internal dependency
• explicitly include ModelDB in RPC endpoints

v0.13.19 (2020-01-08)

New Features

• enable overwriting code and dataset versions
• unpack tarballs in run.fetch_artifacts()

Bug Fixes

• include virtual environment-like directories when automatically logging custom modules

v0.13.18 (2019-12-12)

New Features

• add run.clone()
• add a decorator for models’ predict() to handle argument unpacking

Bug Fixes

• properly propagate deployment error messages
• enable calling run.deploy() and run.undeploy() even if the run is already deployed / not deployed
• properly handle Python 2 string types in querying methods

v0.13.17 (2019-12-05)

Deprecations

• utils.TFSavedModel, in favor of the class-as-model system

New Features

• enable passing more datatypes into DeployedModel.predict()
• add overwrite flag to most artifact logging functions
• enable deployment through ExperimentRun objects
• add a decorator for models’ predict() to handle datatype conversion
• only default to https for endpoints hosted by Verta

Internal Changes

• remove external dependency on six

v0.13.16 (2019-12-02)

New Features

• enable logging directories as ZIP archives with log_artifact()

v0.13.15 (2019-11-27)

New Features

• support logging classes as models
• support associating artifact dependencies with class models
• enable downloading artifacts into a local cache for use with class models

v0.13.14 (2019-11-19)

New Features

• enable indefinite retries on prediction 404s

v0.13.13 (2019-11-18)

Backwards Incompatibilities

• ExperimentRun.log_model() now no longer accepts a user-defined key, and is intended for deployment

Deprecations

• ExperimentRun.log_model_for_deployment(), in favor of more modular logging functions

New Features

• implement ExperimentRun.log_requirements()
• implement ExperimentRun.log_training_data()
• make prediction token optional in DeployedModel::from_url()

Bug Fixes

• retry predictions on non-model 502s

v0.13.12 (2019-11-07)

New Features

• enable indefinite retries on prediction 429s

Bug Fixes

• accommodate external 502s on predictions

Internal Changes

• pass host URL scheme to back end
• reduce dataset version name collisions in tests

v0.13.11 (2019-10-30)

Backwards Incompatibilities

• slightly bump dependency versions and remove grpcio

Bug Fixes

• obtain DatasetVersion timestamps robustly for Python 2

Internal Changes

• clean up Datasets generated during tests
• skip tests on missing imports instead of failing

v0.13.10 (2019-10-27)

Bug Fixes

• fix bug with locally-hosted artifact stores

Internal Changes

• update notebooks

v0.13.9 (2019-10-17)

Bug Fixes

• replace json.JSONDecodeError for Python 2
• remove check for Verta credentials from DeployedModel::from_url()
• properly resolve relative paths in deployment for custom modules
• enable uploading non-Python artifacts
• enable consistent retrieval of models for Python 2

Internal Changes

• add retries for HTTP 502s

v0.13.8 (2019-10-03)

New Features

• enable logging a setup script for the beginning of model deployment
• add verta to uploaded requirements if not present

Internal Changes

• revise pytests

v0.13.7 (2019-09-18)

New Features

• accept key prefixes for S3DatasetVersion
• implement verta.deployment.DeployedModel

Bug Fixes

• enable code version to be downloaded as a ZIP archive through the Web App
• fix bug in run.get_dataset_version()
• fix bug in dataset.get_latest_version()
• catch all unpickling-related errors in get_artifact()

Internal Changes

• keep cell execution numbers in example notebooks

v0.13.6 (2019-09-05)

Bug Fixes

• fix small bugs in the _dataset submodule

Internal Changes

• update protos

v0.13.5 (2019-09-05)

Bug Fixes

• fix various bugs in the _dataset submodule

v0.13.3 (2019-09-04)

Deprecations

• client.expt_runs, because its meaning is ambiguous; proj.expt_runs and expt.expt_runs are preferred
• ret_all_info parameter in querying methods, because it returns user-unfriendly objects

New Features

• implement client.set_experiment_run(id=…)
• implement dataset retrieval functions
• propagate error messages from the back end

Bug Fixes

• support run.get_*() when the value is None
• fix bug where Project, Experiment, and ExperimentRun objects couldn’t be pickled
• fix bug when Datasets are created in Python 2
• log DatasetVersion timestamps as milliseconds, as expected by the Web App
• fix bug when the working directory is captured by run.log_modules()
• fix bug when run.log_modules() is used in Python 2
• fix bug when querying methods are called from an empty ExperimentRuns
• perform basic key validation in querying methods

Internal Changes

• create testing fixtures for deterministic input spaces
• fix data versioning tests
• fix non-artifact tests
• fix artifact tests
• implement model logging tests
• implement basic querying method tests

v0.13.2 (2019-08-20)

New Features

• add ExperimentRun.get_dataset_version()

v0.13.1 (2019-08-20)

Bug Fixes

• handle more states in DatasetVersion.__repr__()

v0.13.0 (2019-08-20)

New Features

• enable file extensions on artifacts in the Web App
• support basic data versioning

Bug Fixes

• convert everything to new-style classes for Python 2 compatibility

Internal Changes

• support dynamically fetching custom deployment URLs
• make Pillow an optional dependency
• support potentially handling a 401 on verifyConnection

v0.12.9 (2019-08-13)

New Features

• support passing in a full URL as the host parameter to Client()

Bug Fixes

• fix bugs regarding logging and retrieving datasets

Internal Changes

• propagate more deployment errors to the Client

v0.12.8 (2019-08-08)

Internal Changes

• bump patch version to 8, to celebrate August 8th
• handle getting Verta environment variables more consistently

v0.12.7 (2019-08-08)

New Features

• support logging functions for deployment
• ignore virtual environment directories when logging custom modules for deployment

Bug Fixes

• define source code UTF-8 encoding for Python 2 compatibility
• use new-style classes for Python 2 compatibility

Internal Changes

• implement DeployedModel::from_url() factory method
• propagate runtime errors to the Client during DeployedModel.predict()
• add custom module logging example notebook

v0.12.6 (2019-08-01)

New Features

• implement a compress parameter on demo predict utility to enable request body compression

Internal Changes

• reduce redundancies in demo predict utility

v0.12.5 (2019-07-26)

New Features

• implement a debug parameter and attribute on Client to print verbose debugging information

v0.12.4 (2019-07-25)

New Features

• remove the need for log_modules()’s second argument (search_path)

v0.12.3 (2019-07-17)

Bug Fixes

• ensure ModelAPI value names are cast to str

Internal Changes

• identify model types by superclass
• update example notebooks with proper ModelAPI instantiation
• update demo notebook with log_code()

v0.12.2 (2019-07-16)

Bug Fixes

• move Git repo check from Client init to log_code()

v0.12.1 (2019-07-16)

Backwards Incompatibilities

• The non-public prediction utility now uses our updated REST prediction endpoint

New Features

• implement log_code() and get_code() for code versioning
• allow periods in Artifact get functions
• enable retrieving integers as integers (instead of as floats) from the back end

Bug Fixes

• catch and raise duplicate column name error on ModelAPI initialization
• properly handle daylight saving time when logging observation timestamps

Internal Changes

• implement internal Configuration utility struct
• add PyTorch example notebook
• implement internal utility for unwrapping directory paths into contained filepaths
• implement internal utilities for reading Git information from the local filesystem
• implement internal utilities for finding executing Python source files
• implement internal utility for getting the file extension from a filepath
• log file extensions with Artifacts

v0.12.0 (2019-06-27)

Backwards Incompatibilities

• The dump() and load() functions have been removed from the public utils module.

New Features

• implement ignore_conn_err parameter and attribute to Client
• implement log_modules() for uploading custom Python modules for deployment

Bug Fixes

• enable logging lists, and dictionaries with string keys, as attributes on client.set_*() to match run.log_attribute()
• simplify stack traces by suppressing contexts during handling for a remaining handful of raise statements
• add missing error message to get_observation()

Internal Changes

• use internal Connection utility object for connection configuration
• define Artifact Store bucket names using a checksum of the artifact
• check for dataset CSV existence before wget in census-end-to-end.ipynb
• expand and unify gitignores

v0.11.7 (2019-06-10)

Backwards Incompatibilities

• The constructors for Project, Experiment, ExperimentRun, and ExperimentRuns—as well as with their _get() and _create() functions—now take an additional retry parameter, though these functions are all not intended for public use to begin with.

New Features

• enable logging lists, and dictionaries with string keys, as attributes
• implement a max_retries parameter and attribute on Client to retry requests with exponential backoff on 403s, 503s, and 504s

Internal Changes

• delegate most REST calls to an internal utility function
• implement back end load test
• change Read the Docs sidebar from fixed to static
• fix a bug that matplotlib has with macOS which was restricting testing

v0.11.6 (2019-06-07)

Backwards Incompatibilities

• Providing a cloudpickle version in the requirements for deployment that doesn’t match the version used by the Client now raises an error instead of overwriting the line in the requirements.

New Features

• add ExperimentRun’s Verta WebApp URL to its __repr__()

Bug Fixes

• use cloudpickle.__version__ instead of relying on pip

Internal Changes

• remove internal utility get_env_dependencies()
• update notebooks

v0.11.5 (2019-06-04)

Backwards Incompatibilities

• The dataset_csv parameter for log_model_for_deployment() has been replaced with two parameters for feature and target DataFrames.

Bug Fixes

• properly render lists in docstrings

Internal Changes

• have the upload script clean out build directories after uploading

v0.11.4 (2019-05-31)

Backwards Incompatibilities

• The dataset_df parameter for log_model_for_deployment() has been renamed to dataset_csv.

Bug Fixes

• reset the correct streams in log_model_for_deployment() instead of model_api over and over again

v0.11.3 (2019-05-31)

New Features

• implement __version__ attribute on package

Bug Fixes

• remove unsupported dependency on pandas and NumPy in utils module

Internal Changes

• move package version string from verta/setup.py to verta/verta/__about__.py
• remove old model API tests that have been superseded by property-based tests
• add pandas as a testing dependency

v0.11.2 (2019-05-30)

Backwards Incompatibilities

• Parameters for client.set_* functions have been renamed to name and id, from e.g. proj_name and _proj_id.
• The _id attribute of Project, Experiment, and ExperimentRun have been renamed to id.
• The default generated names for Project, Experiment, and ExperimentRun have been shortened.

Bug Fixes

• fix typos in client.set_* error messages

v0.11.1 (2019-05-29)

Bug Fixes

• fix internal utility get_env_dependencies() for compatibility with Python 3.6 and earlier

v0.11.0 (2019-05-29)

Backwards Incompatibilities

• log_model_for_deployment() now no longer requires a dataset argument, but requires a model API argument. The order of parameters has changed, and dataset_csv has been renamed to dataset_df.

New Features

• implement ModelAPI utility class for generating model APIs

Internal Changes

• create an example notebook that downloads our beloved Census data with wget
• rename the “scikit” model type to “sklearn”
• delete old internal model API generation utility
• update demo utility predict function to simply dump the JSON input into the request body
• implement internal utility to check for exact version pins in a requirements.txt
• implement internal utility to obtain the local environment’s Python version number
• update READMEs
• add utils module to API reference
• implement tests for model API generation
• implement property-based tests for model API generation
• add deepdiff to testing requirements
• add hypothesis to testing requirements

v0.10.2 (2019-05-22)

no functional changes

v0.10.1 (2019-05-22)

Bug Fixes

• properly expose intermediate subpackages for compatibility with Python 3.2 and earlier

v0.10.0 (2019-05-16)

Backwards Incompatibilities

• log_hyperparameters() now must take a single, whole dictionary as an argument and no longer accepts dictionary unpacking.
• Getting observations from an ExperimentRun now returns tuples pairing observations with their timestamps.
• Passing a string into artifact logging functions now attempts to open a file located at the path represented by that string, rather than simply logging the string itself.
• Attempting to log an unsupported datatype now throws a TypeError instead of a ValueError.
• Logging artifacts now uses cloudpickle by default, instead of pickle.
• The internal logic for getting a Project by name has changed, and will be incompatible with old versions of the Verta Back End.
• The internal logic for handling uploading custom models for deployment has changed, and will be incompatible with old versions of the Verta Back End.
• The internal logic for getting an ExperimentRun by name has changed, and may be incompatible with old versions of the Verta Back End.

New Features

• associate user-specified or automatically-generated timestamps with observations
• implement methods on ExperimentRun for logging and getting tags
• implement methods on ExperimentRun for logging multiple attributes, metrics, or hyperparameters in a single transaction
• enable uploading custom model APIs for deployment
• create functions specifically for logging artifact paths without attempting uploads

Bug Fixes

• reset stream pointer on failed deserialization attempts

Internal Changes

• convert pandas DataFrames into CSVs when logging for deployment for data monitoring
• implement a secondary predict function in demo utilities that returns the raw HTML response instead of a formatted response
• move our example notebooks from workflows/demos/ to workflows/examples/
• change “unknown” model type to “custom” in model API
• add “keras” deserialization in model API
• add cloudpickle to requirements with the locally pinned version if it was used when logging for deployment
• implement handful of small fixes to maintain Python 2.7 compatibility