PyCon 2013 tutorial: parallel ML before it was mainstream

What it does A set of executable IPython notebooks from a 2013 PyCon tutorial by Olivier Grisel, covering how to parallelize scikit-learn workflows across cores and cheap EC2 spot instances. Topics span cross-validation, grid search, text feature extraction, memory-mapped numpy arrays, and spinning up clusters with the since-deprecated StarCluster tool.

The interesting bit The tutorial captures a specific evolutionary moment: scikit-learn’s Estimator API was still being learned, IPython (not yet Jupyter) was the interactive frontier, and “cheap parallel compute” meant wrangling StarCluster on Amazon spot instances. The material is adapted from a SciPy 2013 tutorial by Gael Varoquaux and Jake VanderPlas.

Key highlights

• Static rendered notebooks viewable without installation via nbviewer.ipython.org
• Covers numpy memory mapping for node-level memory optimization
• Includes fetch_data.py to pull datasets before running interactively
• Explicitly targets developers already comfortable with scikit-learn basics
• Video recording available for following along with notebook titles as section markers

Caveats

• Setup instructions reference IPython 2.2.0 and scikit-learn 0.15.2; modern environments will need translation
• StarCluster dependency for EC2 clustering is unmaintained (last release 2013)
• No updates since original publication; some APIs have evolved significantly

Verdict Worth browsing for historical context on how the scikit-learn ecosystem taught parallelism, or if you’re maintaining legacy IPython-based workflows. Skip if you need current best practices—modern Dask, Ray, or scikit-learn’s own n_jobs patterns have superseded much of this.