CLOUDFLARE’S Workflows control plane has been rearchitected to handle a much higher throughput, moving from a single Durable Object bottleneck to a distributed, scalable design. The new V2 architecture introduces two components, SousChef and Gatekeeper, to share metadata and manage concurrency across multiple workflows within an account, while ensuring engines never overwhelm the account.
As a result, developers can run up to 50,000 concurrent instances, up from 4,500, and create 300 instances per second per account, up from 100, with up to 2 million queued instances per workflow, up from 1 million. The migration from V1 to V2 involved converting account-level Durable Objects into SousChef-like objects and rolling out the new Account DO alongside the SousChefs, enabling a seamless, low-downtime transition for customers.
V1’s bottlenecks—where the Account DO acted as a central registry for all workflows in an account—were addressed by distributing responsibility and introducing a periodic, fair queuing mechanism via Gatekeeper. The redesign also realigns operation paths so that each instance check, metadata fetch, and storage step happens with minimal latency, supporting larger-scale agent-driven workflows.