Dask

When your pandas DataFrame won't fit in RAM or NumPy operations are crawling on a single core, this is what you reach for. It parallelizes familiar pandas and NumPy APIs across multiple cores or machines without rewriting your code. The lazy evaluation model is key: operations build a task graph until you call compute(), so you can chain transformations efficiently. Choose DataFrames for tabular data, Arrays for scientific computing, Bags for log files and JSON, or Futures when you need fine-grained control over task execution. The threaded scheduler works great for numeric operations, but switch to processes for Python-heavy work that hits the GIL.