Async Python Is Secretly Deterministic

The article from DBOS explores how their team leveraged a nuanced aspect of Python's asyncio event loop to build deterministic durable workflows, which are crucial for replay-based recovery. It delves into the internal workings of asyncio to reveal a "secret" determinism that, at first glance, might seem counterintuitive for concurrent programming.

• Durable execution systems require deterministic workflows for replay-based recovery, which traditionally poses a challenge with concurrent async operations where task completion order can vary.
• While asyncio.gather is commonly used for running concurrent tasks, the exact interleaving of their execution can be non-obvious and seemingly non-deterministic.
• At its core, asyncio operates on a single-threaded event loop that uses a scheduler to process a queue of tasks.
• Crucially, when new tasks are created and scheduled (e.g., via asyncio.gather), the event loop processes them in a strict First-In, First-Out (FIFO) order.
• Although the completion order and subsequent interleaving of these tasks can be unpredictable, the initial order in which they begin execution is deterministic.
• DBOS exploits this by assigning unique step IDs before the first await call within a task, ensuring that workflow steps are processed in a consistent, predictable order.
• This deep understanding of the asyncio event loop allows for writing concurrent code that is both deterministic and inherently safer, simplifying the reasoning behind complex async systems.

The author concludes that appreciating these subtleties of the asyncio event loop empowers developers to create more reliable and predictable systems, inviting readers to explore DBOS's tools for durable workflows.