Streaming execution for performance

The Polars lazy API can execute queries in streaming mode, processing data in batches instead of all at once, so datasets that do not fit in memory can still be processed; the streaming engine is more performant than the in-memory engine. In Python, pass engine="streaming" to collect; in Rust, use with_streaming(true). Some operations are inherently non-streaming or not implemented in a streaming manner yet, and Polars falls back to the in-memory engine for those; a user does not have to know about this, but it can be interesting for debugging memory or performance issues. Plot the physical graph via show_graph with plan_stage="physical"; its legend shows memory intensity. The streaming engine can report per-node runtime metrics to Polars Cloud to uncover bottlenecks. Scope: Polars. Reviewed October 4, 2026.