Windowing

Overview

In the world of data processing on an unbounded stream, Windowing is a concept of grouping data using temporal boundaries. We use event-time to discover temporal boundaries on an unbounded, infinite stream and Watermark to ensure the datasets within the boundaries are complete. The reduce is applied on these grouped datasets. For example, when we say, we want to find number of users online per minute, we use windowing to group the users into one minute buckets.

The entirety of windowing is under the groupBy section.

vertices:
  - name: my-udf
    udf:
      groupBy:
        window: ...
        keyed: ...

Since a window can be Non-Keyed v/s Keyed, we have an explicit field called keyedto differentiate between both (see below).

Under the window section we will define different types of windows.

Window Types

Numaflow supports the following types of windows

• Fixed
• Sliding
• Session

Non-Keyed v/s Keyed Windows

Non-Keyed

A non-keyed partition is a partition where the window is the boundary condition. Data processing on a non-keyed partition cannot be scaled horizontally because only one partition exists.

A non-keyed partition is usually used after aggregation and is hardly seen at the head section of any data processing pipeline. (There is a concept called Global Window where there is no windowing, but let us table that for later).

Keyed

A keyed partition is a partition where the partition boundary is a composite key of both the window and the key from the payload (e.g., GROUP BY country, where country names are the keys). Each smaller partition now has a complete set of datasets for that key and boundary. The subdivision of dividing a huge window-based partition into smaller partitions by adding keys along with the window will help us horizontally scale the distribution.

Keyed partitions are heavily used to aggregate data and are frequently seen throughout the processing pipeline. We could also convert a non-keyed problem to a set of keyed problems and apply a non-keyed function at the end. This will help solve the original problem in a scalable manner without affecting the result's completeness and/or accuracy.

When a keyed window is used, an optional partitions can be specified in the vertex for parallel processing.

Usage

Numaflow supports both Keyed and Non-Keyed windows. We set keyed to either true (keyed) or false (non-keyed). Please note that the non-keyed windows are not horizontally scalable as mentioned above.

vertices:
  - name: my-reduce
    partitions: 5 # Optional, defaults to 1
    udf:
      groupBy:
        window: ...
        keyed: true # Optional, defaults to false