Airflow vs MLflow: What are the differences?
Airflow: A platform to programmaticaly author, schedule and monitor data pipelines, by Airbnb. Use Airflow to author workflows as directed acyclic graphs (DAGs) of tasks. The Airflow scheduler executes your tasks on an array of workers while following the specified dependencies. Rich command lines utilities makes performing complex surgeries on DAGs a snap. The rich user interface makes it easy to visualize pipelines running in production, monitor progress and troubleshoot issues when needed; MLflow: An open source machine learning platform. MLflow is an open source platform for managing the end-to-end machine learning lifecycle.
Airflow and MLflow are primarily classified as "Workflow Manager" and "Machine Learning" tools respectively.
Some of the features offered by Airflow are:
- Dynamic: Airflow pipelines are configuration as code (Python), allowing for dynamic pipeline generation. This allows for writting code that instantiate pipelines dynamically.
- Extensible: Easily define your own operators, executors and extend the library so that it fits the level of abstraction that suits your environment.
- Elegant: Airflow pipelines are lean and explicit. Parameterizing your scripts is built in the core of Airflow using powerful Jinja templating engine.
On the other hand, MLflow provides the following key features:
- Track experiments to record and compare parameters and results
- Package ML code in a reusable, reproducible form in order to share with other data scientists or transfer to production
- Manage and deploy models from a variety of ML libraries to a variety of model serving and inference platforms
Airflow and MLflow are both open source tools. It seems that Airflow with 13.3K GitHub stars and 4.91K forks on GitHub has more adoption than MLflow with 30 GitHub stars and 13 GitHub forks.
I am so confused. I need a tool that will allow me to go to about 10 different URLs to get a list of objects. Those object lists will be hundreds or thousands in length. I then need to get detailed data lists about each object. Those detailed data lists can have hundreds of elements that could be map/reduced somehow. My batch process dies sometimes halfway through which means hours of processing gone, i.e. time wasted. I need something like a directed graph that will keep results of successful data collection and allow me either pragmatically or manually to retry the failed ones some way (0 - forever) times. I want it to then process all the ones that have succeeded or been effectively ignored and load the data store with the aggregation of some couple thousand data-points. I know hitting this many endpoints is not a good practice but I can't put collectors on all the endpoints or anything like that. It is pretty much the only way to get the data.
For a non-streaming approach:
You could consider using more checkpoints throughout your spark jobs. Furthermore, you could consider separating your workload into multiple jobs with an intermittent data store (suggesting cassandra or you may choose based on your choice and availability) to store results , perform aggregations and store results of those.
Spark Job 1 - Fetch Data From 10 URLs and store data and metadata in a data store (cassandra) Spark Job 2..n - Check data store for unprocessed items and continue the aggregation
Alternatively for a streaming approach: Treating your data as stream might be useful also. Spark Streaming allows you to utilize a checkpoint interval - https://spark.apache.org/docs/latest/streaming-programming-guide.html#checkpointing