Building Robust Pipelines with Airflow
1 like1,977 views
The document discusses Zymergen's approach to building robust experimentation pipelines using Airflow for managing complex workflows in genetic engineering and fermentation processes. It highlights challenges such as process failures, biases in data, and the integration of diverse data sources through custom ETL solutions. The use of tools like Airflow and Bayesian models is emphasized for flexible and modular workflow management in machine learning and experimental setups.
Building Robust Pipelines with Airflow
- 1. @erinshellman Wrangle Conf July 20th, 2017 Building Robust Pipelines with Airflow
- 2. Zymology: is the science of fermentation and it’s applied to make materials and molecules ! " # $ Beer Insulin Food additives Plastics
- 4. Zymergen provides a platform for rapid improvement of microbial strains through genetic engineering.
- 5. Robotic automation Our experimentation is increasingly orchestrated with robotics and machine learning.
- 6. Learning how to efficiently navigate the genome is the mission of data science at Zymergen
- 7. Blocker: process failure Orchestrating complex experiments with robots is hard, and there are process failures. These failures often cause sporadic, extreme measurement values.
- 8. Blocker: batch effects We see temporal effects based on when experiments were performed
- 9. Blocker: different interpretations of results We’re building a platform that can support any microbe and any molecule. Sometimes that results in a proliferation of solutions with disagreement on which is best.
- 10. Processing pipeline 1.Identify process failures 2.Quantify and remove process- related bias 3.Identify strains that show improvement using consistent criteria Clean model inputs Outlier detection Normalization Hit detection
- 11. Rolling our own ETL pipeline There are many ways to measure the concentration of a molecule. Any microbe, any molecule… any experiment, many data formats.
- 12. Describing complex processing dependencies is hard. Rolling our own ETL pipeline
- 13. Airflow https://airflow.incubator.apache.org/ “Airflow is a platform to programmatically author, schedule and monitor workflows.” Airflow gives us flexibility to apply a common set of processing steps to variable data inputs, schedule complex processing workflows, and has become a delivery mechanism for our products.
- 15. e.g. Normalization Airflow workflows are described as directed acyclic graphs (DAGs). Each task node in the DAG is an operator.
- 16. The anatomy of a DAG Custom operators Ordering Instantiate DAG
- 18. Airflow + PyStan With Bayesian hierarchical models we estimate (and monitor) the distribution of batch effects. Experimental bias
- 19. DropBox • Scientists at Zymergen work with data using many different tools including JMP, SQL, and Excel. • We use a custom DropBox hook to make quick data ingestion pipelines.
- 21. 3rd-party hooks & operators
- 22. Operator
- 23. Pairs well with Superset! “Apache Superset is a modern, enterprise-ready business intelligence web application” https://github.com/apache/incubator-superset
- 25. Fairflow: Functional Airflow • The core of Fairflow is an abstract base class foperator that takes care of instantiating your Airflow operators and setting their dependencies. • In Fairflow, DAGs are constructed from foperators that create the upstream operators when the final foperator is called.
- 26. Configuring complex ML workflows… functionally
- 27. Defining ML workflows In the DAG definition, create an instance of the task. Then, instantiate a DAG like usual and call the compare task on the DAG.
- 28. Defining ML workflows The design allows for simple creation of complicated experimental workflows with arbitrary sets of models, parameters, and evaluation metrics.
- 29. Is Airflow for you? Do you have heterogeneous data sources? Do you have complex dependencies between processing tasks? Do you have data with different velocities? Do you have constraints on your time? Probably!