The WAS takes the summary of execution of a workflow and uses a random forest algorithm to analyse the workflow execution and suggest improvements to the workflow. The WAS handles both classification and regression when predicting labels.
The WAS takes a JSON file as input. The accepted format of the file is quite flexible, however the following keys are expected:
A workflow key, which in turn contains a name key, respectively providing the workflow's information, and the
workflow's name.
workflow also needs to contain a nodes key. nodes contains a flexible number of workflow nodes/tasks.
Each node is expected to have an id and a name.
If any node has successor nodes, the successors are expected to be stored by id under a successor-id key.
The successor-id key can be flexible in nomenclature as long as the words successor and id are contained in the
key.
The input JSON is iterated through each workflow and each node. If any of the nodes is a subworkflow, it should contain
the key subWorkflow.
If the nodes contain nested dictionaries, they will be flattened. This means that if a node value is given as:
"nodes": [{"executionStatistics": {"executionDurationSinceStart": 811} } ]
the value will be saved as:
executionStatistics.executionDurationSinceStart = 811.
The WAS expects and accepts a number of arguments when called.
-f or --file passing the file-path to the JSON input file. Required.
-u or --user passing the username of the user executing the WAS. Required
-tc or --task_classifier passing one or multiple nodes/tasks labels for the classifier. The labels should be
categorical values. The labels must exists in the JSON file. Optional.
-tr or --task_regressor passing one or multiple nodes/tasks labels for the regressor. The labels must be continuous
numerical values. The labels must exists in the JSON file. Optional.
-wc or --workflow_classifier passing one or multiple workflow labels for the classifier. The labels should be
categorical values. The labels must exists in the JSON file. Optional.
-wr or --workflow_regressor passing one or multiple workflow labels for the regressor. The labels must be continuous
numerical values. The labels must exists in the JSON file. Optional.
-tf or --task_features passing names of keys (remember to account for the flattening of the input) which are to be
considered as features by the machine learning when attempting to predict the task labels. If none are given, then all keys given by
the JSON file are used. Optional.
-wf or --workflow_features passing names of keys (remember to account for the flattening of the input) which are to
be considered as features by the machine learning when attempting to predict the workflow labels. If none are given, then all keys
given by the JSON file are used. Optional.
The above command line arguments can be provided in the form of a profile. Profiles are text files containing the
desired parameters. The file can be passed when calling the WAS using the @ symbol, for example:
python WorkflowAnalysisSystem.py @../profiles/user_2
Example profiles are provided in the 'profiles' directory.
The WAS uses the random forest predictions to analyse the labels and features of the workflow execution. For each of the workflows and subworkflows provided as input, as well as for each task, the ML attempts to predict the labels and returns both the predictions, and the features it deemed to be most important to infer the label value.
The ML predicts using any historical data available which matches the current workflows and tasks.
If the prediction error of the ML for a given label is too high (over 20%), no important features are returned as the ML inference cannot be trusted.
To aid the user in improving the workflow, a topological analysis is also performed on the execution data.
The topological analysis first analyses any paths that branch out of a single node/task and reunite on a secondary node/task. For each label and important feature, it calculates the ratios (if a categorical values), or the mean (if a continous numerical values) of the branch's tasks using the historical data available.
Then, for each task provided by the JSON input it analysis the historical performance, and returns the calculated ratios (if a categorical values) or mean (if a continous numerical values).
The WAS outputs a report containing all the above-detailed analyses in xlsx format.
To install all the required packages, navigate to the src directory and run:
pip install -r requirements.txt
Otherwise, the following packages must be manually installed:
pandas:
pip install pandas
matplotlib:
pip install matplotlib
scikit-learn:
pip install scikit-learn
Jinja2:
pip install Jinja2
xlsxwriter:
pip install xlsxwriter
Python 3.6 or newer must also be installed.
To run the tests available in the tests dir you need to install pytest using:
pip install pytest
Then, from the 'src' dir, run:
python -m pytest