Data CI

DataCI is a comprehensive platform for track your data-centric AI pipeline in dynamic streaming data settings. It provides rich APIs for seamless streaming dataset management, data-centric pipeline development, versioning and evaluation on streaming scenarios.

DataCI is featured for

• :hatching_chick: Easy to use: Enable DataCI management by only few lines of modification to your existing data-centric functions and pipelines
• :card_index_dividers: Version Control: Track versions of streaming data, data-centric pipelines, running results and their lineages
• :ocean: Streaming Simulation: Testbed for developing and evaluating data-centric pipelines under simulated streaming settings (WIP)

This project is still under development. Please feel free to open an issue if you have any questions or suggestions.

• Installation
• Quick Start
• Examples (for Advanced Usages)
• Citation
• Contributors
• License

Installation :construction:

Easy install DataCI from PyPI

pip install ml-data-ci

Manual install DataCI package from source code:

pip install .

:warning: For pipeline orchestration backend framework (i.e., Apache-Airflow), you may need to install manually since the version to meet your deployment can be different. Here is a snapshot for quick installation:

AIRFLOW_VERSION=2.6.1
PYTHON_VERSION="$(python --version | cut -d " " -f 2 | cut -d "." -f 1-2)"
CONSTRAINT_URL="https://raw.githubusercontent.com/apache/airflow/constraints-${AIRFLOW_VERSION}/constraints-${PYTHON_VERSION}.txt"
pip install "apache-airflow[celery]==${AIRFLOW_VERSION}" --constraint "${CONSTRAINT_URL}"

Initialize DataCI by:

dataci init

:warning: This command will initialize the DataCI, you should run this command only once when installing.

Start DataCI server by:

dataci start

Quick Start :truck:

With few lines of modification, you can easily track every version of the input, output data and code for your existing data pipeline. You can also see the below example script here :scroll:.

1. Write a data augmentation stage:

   import augly.text as textaugs
   from dataci.plugins.decorators import stage
   
   
   @stage
   def text_augmentation(df):
       aug_function = textaugs.ReplaceSimilarUnicodeChars()
       df['text'] = aug_function(df['text'].tolist())
       return df
   

2. Define the text processing CI pipeline:

   from datetime import datetime
   from dataci.models import Event
   from dataci.plugins.decorators import dag, Dataset
   
   
   @dag(
       start_date=datetime(2020, 7, 30), schedule=None,
       trigger=[Event('publish', 'yelp_review@*', producer_type='dataset', status='success')],
   )
   def text_process_ci():
       # Instead of manually track the dataset version: 
       #     raw_dataset_train = df.read_csv(f'yelp_review_{DATASET_VERSION}/train.csv')
       # input dataset versioning is automatically handled by DataCI
       raw_dataset_train = Dataset.get('yelp_review@latest')
       text_aug_df = text_augmentation(raw_dataset_train)
   
       # Instead of manually track the output dataset version:
       #    text_aug_df.to_csv(f'text_aug_{DATASET_VERSION}/train.csv')
       # output dataset versioning is automatically handled by DataCI.
       Dataset(name='text_aug', dataset_files=text_aug_df)
   
   # Build the pipeline
   text_process_ci_pipeline = text_process_ci()
   

   We write the workflow process in a function text_process_ci in a normal way, and then decorate it with @dag to convert it to a DataCI tracked workflow. With the trigger argument, we can specify the trigger event for the pipeline. In this example, the pipeline will be triggered when a new version of yelp_review dataset is published.

3. Manually run the pipeline:

   from dataci.models import Dataset
   
   
   if __name__ == '__main__':
       # Prepare the input dataset, you can either provide a list or a file path
       yelp_review_dataset = Dataset('yelp_review', dataset_files=[
           {'date': '2020-10-05 00:44:08', 'review_id': 'HWRpzNHPqjA4pxN5863QUA', 'stars': 5.0, 'text': "I called Anytime on Friday afternoon about the number pad lock on my front door. After several questions, the gentleman asked me if I had changed the battery.",},
           {'date': '2020-10-15 04:34:49', 'review_id': '01plHaNGM92IT0LLcHjovQ', 'stars': 5.0, 'text': "Friend took me for lunch.  Ordered the Chicken Pecan Tart although it was like a piece quiche, was absolutely delicious!",},
           {'date': '2020-10-17 06:58:09', 'review_id': '7CDDSuzoxTr4H5N4lOi9zw', 'stars': 4.0, 'text': "I love coming here for my fruit and vegetables. It is always fresh and a great variety. The bags of already diced veggies are a huge time saver.",},
       ]).publish(version_tag='2020-10')
       # Test the pipeline locally
       text_process_ci_pipeline.test()
       # Publish the pipeline: text_process_ci@v1 and its stage text_augmentation@v1 will be tracked
       text_process_ci_pipeline.publish()
       # Run the pipeline on the server
       run_id = text_process_ci_pipeline.run()
   

   The text process CI pipeline will be triggered and run. Go to pipeline runs dashboard to check the pipeline run result.
   Note that the you should write pipeline testing and trigger code aside from the pipeline definition code. Otherwise, the pipeline will be triggered every time when you import the pipeline definition code.

4. Automatically trigger the CI pipeline by publish a new version of input dataset:

   from dataci.models import Dataset
   
   
   if __name__ == '__main__':
       # DataCI auto track the new version of yelp_review dataset
       yelp_review_dataset = Dataset('yelp_review', dataset_files=[
           {'date': '2020-11-02 00:59:29', 'review_id': 'LTr95e6eOmLc7S_1WxM88Q', 'stars': 5.0, 'text': "First of all the owner and staff went above and beyond to make us feel comfortable and protected during covid.  Secondly, the bar had fantastic drinks",},
           {'date': '2020-11-16 22:07:48', 'review_id': 'UhA9H0LK59qZegWOxyotcA', 'stars': 5.0, 'text': "Herbies is awesome! My brunch was perfect, service, food, and drinks! I will definitely continue coming back.",},
           {'date': '2020-11-18 06:38:46', 'review_id': 'YqTMxlbebNBDcKYTIUvrdw', 'stars': 5.0, 'text': "You won't regret stopping here, hidden gem with great food and a laid back and comfortable atmosphere, a place you can gather with friends and they will treat you like family while you're there.",},
       ]).publish(version_tag='2020-11')
   

   Before executing the above code, please wait some time for the previous manual triggered run to finish. Upon the publish of the new version of input dataset, the text process CI pipeline is automatically triggered and evaluate on the new version of streaming data. Go to pipeline runs dashboard to check the pipeline run result.

Examples :books:

We have provided several examples to demonstrate the advanced usage of DataCI:

• Continuous Text Classification Development with Streaming Data
• DataCI Playground (WIP)

Citation :seedling:

Our tech report is available at arXiv:2306.15538 and DMLR@ICML'23. Please cite as:

@article{zhang2023dataci,
   title = {DataCI: A Platform for Data-Centric AI on Streaming Data},
   author = {Zhang, Huaizheng and Huang, Yizheng and Li, Yuanming},
   journal = {arXiv preprint arXiv:2306.15538},
   year = {2023}
}

Contributors :sparkles:

• Yuanming Li
• Huaizheng Zhang
• Yizheng Huang

License :page_facing_up:

This repository is open-sourced under MIT License.