Airbnb Open-Sources Skipper Durable Workflow Engine for JVM Services

Airbnb open-sourced Skipper, a lightweight durable workflow execution library for Java and Kotlin services that embeds in existing services and checkpoints workflow state to the existing database, eliminating the need for a separate orchestration cluster. Multiple service replicas coordinate tasks via leases, and recovery replays workflow methods while returning persisted action results. Skipper does not guarantee exactly-once downstream execution, so developers must ensure idempotency; the open-source version provides MySQL and SQLite adapters, with Postgres and DynamoDB planned.
Event version 1.
Key insights
Library, not a cluster
Skipper is a library added to an existing service, with no control plane, cluster, or new datastore to operate. Workflow state is checkpointed to the database the service already uses.
Lease-based replica coordination
Multiple service replicas share the database and coordinate tasks using leases. If one replica fails, another can take over.
Replay with persisted results
On recovery, Skipper replays the workflow method from the start, but actions already persisted return their saved results instead of re-executing. Workflows can also pause for signals or timers without holding threads.
No exactly-once guarantee
Skipper cannot guarantee downstream operations occur only once. Some checkpoints are batched by default, and downstream calls and checkpoints are not one atomic transaction, so developers must ensure idempotency for operations such as payments and coupon issuance.
Adapters and internal scale
The open-source version provides MySQL and SQLite adapters, with MySQL recommended for multi-instance production; Postgres and DynamoDB are planned. Airbnb reported internal coverage of more than 15 use cases peaking at 10,000 workflows per second, a figure measured with DynamoDB.