Journey to 1000 Models: Scaling Instagram's algorithm without the Reliability Nightmare
summary
Instagram scales from O(10) GPU models to over 1,000 in production and replaces manual release and monitoring with a model registry that maps numeric IDs to business function and criticality, launch tooling that cuts launches from a week to under a day and supports >10 parallel launches, and a model stability metric built from calibration and normalized entropy to produce SLIs/SLOs for real‑time alerts; capacity moves to virtual pools and one‑click estimations finish within an hour.
AI-written summary · about
chapters (6)
- 0:00Scaling Instagram's recommendations
- 1:50Managing a thousand models
- 3:14Unified model registry at Meta
- 6:25Automating model launches
- 11:38Model health definition & ops
- 16:55Key takeaways & challenges
related talks
18mInstagram in focus
42mImproving Machine Learning Development Reliability
42mModeling Aesthetic Identity- Building a Digital Twin
39mFrom 0 To 500+ Models: Building a Robust Open Sour...
30mWhen Good Models Go Bad: The damage caused by wayward models and how to prevent it
28mComputing Large Foundational Models
31mBeyond One Model: Scaling, Orchestrating & Monitoring
30mOn scaling Machine Learning microservices