Scaling Back-End Deployment
summary
Facebook's unified CD system Conveyor runs ~100,000 deployments per week with a 99.9% success rate by combining scheduled pipelines, automated Push execution, and metric-driven health checks; adoption reached 94% regular deploys after the company-wide "push for push" campaign. Push is refactored into Rust microservices and pluggable push types, adds a translation layer for migration, and cuts bad failures from >1/20 to roughly 1/3,000 while moving 95% of traffic to Push V2.
AI-written summary · about
chapters (8)
- 0:00Conveyor at Facebook
- 2:12Early push tool challenges
- 7:05Low adoption barriers
- 9:32Expanding to all services
- 11:59Scaling Python and monolith
- 13:23Rewriting push in Rust
- 16:12Push V2 with translation
- 17:58Results and future
related talks
24mContinuous Deployment at Scale
23mMoving Fast With Complex...
47mDeploying Like a Boss: From 2 Servers to 20,000
46m10+ Deploys Pe
32mContinuous deployment at scale
01h 03mMoving Fast - without crashing
01h 01mI deploy on Fridays (and maybe you should too)
16mConveyor: One-Tool-Fits-All Continuous Software Deployment
