We present Conveyor, Meta’s software deployment tool,
along with the valuable data obtained from managing over
30,000 deployment pipelines that deploy all kinds of services
at Meta across millions of machines. We describe a wide
range of deployment scenarios that Conveyor supports to
achieve universal coverage. At Meta, out of all the deploy ment pipelines for services deployed via containers, 97% of
them employ fully automated deployments without manual
intervention: 55% utilize continuous deployment, instantly
deploying every code change to production after passing au tomated tests, and the remaining 42% are automatically de ployed on a fixed schedule (mostly daily or weekly) without
manual validation. We highlight several distinguishing fea tures of Conveyor, including safe in-place updates to reduce
hardware costs, analysis of code dependencies to prevent
faulty releases, and the capability to deploy complex ML
models at scale.