Window Operator

The Window Operator implements the Consistent Partition Window relation defined in Substrait.

This operator supports only one window function per instance. To apply multiple window functions, chain multiple Window Operators sequentially.

The operator is stateful and relies on two B+ trees:

• Persistent Tree – Stores all rows in the dataset, ordered by partition columns followed by order-by columns.
• Temporary Partition Tree – Tracks which partitions have been updated (not persisted).

Behavior

When events are received:

• Rows are inserted into the Persistent Tree.
• If a row is deleted (i.e., an event with negative weight), that row's negative output is sent downstream immediately.
• The partition key is calculated and recorded in the Temporary Partition Tree.

For upsert operations, the row is held until a watermark is received, at which point the final calculation is performed.

When a watermark event is received, the operator:

• Iterates over all changed partitions.
• Applies the configured window function.

The logic for how values are calculated within a partition is delegated to the specific window function implementation.

Metrics

The Window Operator exposes the following metrics:

Metric Name	Type	Description
busy	Gauge	Value 0-1 on how busy the operator is.
backpressure	Gauge	Value 0-1 on how much backpressure the operator has.
health	Gauge	Value 0 or 1, if the operator is healthy or not.
events	Counter	How many events the operator outputs.
events_processed	Counter	How many events the operator recieves.