🏃‍♀️ Runhouse Docs 🏠

Why did we build Runhouse? To enable fast, debuggable, and iterable development of machine learning workflows.

Without Runhouse: ML workflow is fragmented and development is slow

• Researchers start in hosted notebooks or SSH’ed into a cluster:

  • Fast and interactive development

  • But usually non-standard compute environment and code

• Research to production happens over the course of days or weeks:

  • Notebook code needs translation to orchestrator nodes

  • Most time spent waiting to rebuild and resubmit pipelines, with each iteration loop taking about 20+ minutes

• Production debugging is challenging:

  • Orchestrators designed for scheduling and logging runs, but are not development-friendly runtimes

  • Continue “debug through deployment” that slowed down research to production in the first place

With Runhouse: Regular code is dispatched to cloud compute for execution at every step

• Researchers write normal code:

  • Each dispatch takes <5 seconds, providing interactive development experience

  • Code executes on the same compute and environment of production

  • Logs stream back to local

• Moving to production is instant:

  • Orchestrator nodes contain 5 lines of dispatch code (not 200 lines of application code)

  • Rather than being your ML runtime, orchestrators are simply used to schedule, log, and monitor runs

• Easily debug or update pipelines in production:

  • Branch the underlying code

  • Make changes and dispatch iteratively, like in the research step

  • Merge back into main

In short, Runhouse makes ML development feel like normal software development.