iiinova — where your digital self keeps evolving

Replies

Maxim Sakhno @maxim-sakhno.bsky.social · 28/01/2026

The logic behind the batches: Kahn’s algorithm for topological sorting. 🧠 LiteStar's implementation: dub.sh/0DG8van LiteStar uses it to identify which dependencies can run in parallel and groups them into batches. The algorithm visualization 👇

A technical diagram illustrating how a Dependency Graph is transformed into execution batches using Kahn's Algorithm. On the left, a graph shows node A depending on B and C, which in turn depend on D, E, and F. An arrow labeled "Kahn's Algorithm" points to the right, showing these nodes organized into three sequential batches. Batch 1 (D, E, F) and Batch 2 (B, C) are contained within boxes labeled "Task Group," indicating parallel execution. Batch 3 contains only node A and is labeled "Await," indicating simple sequential execution.

110

Maxim Sakhno @maxim-sakhno.bsky.social · 28/01/2026

Why use Kahn's algorithm? 🤔 It doesn't just sequence dependencies; it finds batches that can run in parallel. For heavy I/O tasks, this is a massive win. The speedup from concurrency easily outweighs the TaskGroup overhead.

110

Maxim Sakhno @maxim-sakhno.bsky.social · 28/01/2026

But wait... didn't LiteStar beat FastAPI in their own benchmarks? Ah, they used nested dependencies only! That explains it. 💡

A technical benchmark report titled "Dependency injection" comparing FastAPI 0.101.1, LiteStar 2.0.0, and Sanic 23.6.0. The top section lists test cases for resolving 3 nested dependencies in synchronous, asynchronous, and mixed modes. Below, a complex bar chart shows "Requests per second (higher is better)" across these modes. LiteStar significantly outperforms the others in almost every category, particularly in asynchronous dependency resolution, which sets the stage for the specific performance "cliff" discovery detailed in the rest of the thread.

110