SQLite Plus Litestream as the Minimal Stack for Durable Workflows

Obelisk’s author pushes back on the assumption that durable execution demands heavy database infrastructure. Building on DBOS’s “Postgres is all you need” argument, he contends that for many workflow systems the durable surface is just the execution log itself — and a local SQLite file handles that without a network hop, separate control plane, or extra ops burden. Workflows replay from persisted history, activities retry, and compute stays cheap and disposable.

The accumulation problem is solved by Litestream, which asynchronously ships SQLite changes to S3-compatible storage for backup and inspection. The tradeoff is honest: async replication can lose the most recent writes if the volume dies before they ship, so this is not a substitute for a highly available shared database. For bursty, experimental AI agent workloads where each tenant or agent owns a small self-contained state, the pattern shines — a fleet of micro VMs each with their own SQLite file gives better fault isolation than one shared system.

Postgres remains the right call when you need high availability, shared scale, or stronger durability guarantees than async object-storage replication provides. The broader point: most workflow systems should match infrastructure to the state they actually keep, and for the AI agent world that often means SQLite plus Litestream is the sensible default.