The samples on this page demonstrate common custom window patterns. You can create custom windows with [WindowFn
functions]({{ site.baseurl }}/documentation/programming-guide/#provided-windowing-functions). For more information, see the [programming guide section on windowing]({{ site.baseurl }}/documentation/programming-guide/#windowing).
Note: Custom merging windows isn't supported in Python (with fnapi).
You can modify the assignWindows
function to use data-driven gaps, then window incoming data into sessions.
Access the assignWindows
function through WindowFn.AssignContext.element()
. The original, fixed-duration assignWindows
function is:
{% github_sample /apache/beam/blob/master/examples/java/src/main/java/org/apache/beam/examples/snippets/Snippets.java tag:CustomSessionWindow1 %}
To create data-driven gaps, add the following snippets to the assignWindows
function:
For example, the following function assigns each element to a window between the timestamp and gapDuration
:
{% github_sample /apache/beam/blob/master/examples/java/src/main/java/org/apache/beam/examples/snippets/Snippets.java tag:CustomSessionWindow3 %}
Then, set the gapDuration
field in a windowing function:
{% github_sample /apache/beam/blob/master/examples/java/src/main/java/org/apache/beam/examples/snippets/Snippets.java tag:CustomSessionWindow2 %}
After creating data-driven gaps, you can window incoming data into the new, custom sessions.
First, set the session length to the gap duration:
{% github_sample /apache/beam/blob/master/examples/java/src/main/java/org/apache/beam/examples/snippets/Snippets.java tag:CustomSessionWindow4 %}
Lastly, window data into sessions in your pipeline:
{% github_sample /apache/beam/blob/master/examples/java/src/main/java/org/apache/beam/examples/snippets/Snippets.java tag:CustomSessionWindow6 %}
The following test data tallies two users' scores with and without the gap
attribute:
.apply("Create data", Create.timestamped( TimestampedValue.of("{\"user\":\"user-1\",\"score\":\"12\",\"gap\":\"5\"}", new Instant()), TimestampedValue.of("{\"user\":\"user-2\",\"score\":\"4\"}", new Instant()), TimestampedValue.of("{\"user\":\"user-1\",\"score\":\"-3\",\"gap\":\"5\"}", new Instant().plus(2000)), TimestampedValue.of("{\"user\":\"user-1\",\"score\":\"2\",\"gap\":\"5\"}", new Instant().plus(9000)), TimestampedValue.of("{\"user\":\"user-1\",\"score\":\"7\",\"gap\":\"5\"}", new Instant().plus(12000)), TimestampedValue.of("{\"user\":\"user-2\",\"score\":\"10\"}", new Instant().plus(12000))) .withCoder(StringUtf8Coder.of()))
The diagram below visualizes the test data:
![Two sets of data and the standard and dynamic sessions with which the data is windowed.]( {{ “/images/standard-vs-dynamic-sessions.png” | prepend: site.baseurl }})
Standard sessions use the following windows and scores:
user=user-2, score=4, window=[2019-05-26T13:28:49.122Z..2019-05-26T13:28:59.122Z) user=user-1, score=18, window=[2019-05-26T13:28:48.582Z..2019-05-26T13:29:12.774Z) user=user-2, score=10, window=[2019-05-26T13:29:03.367Z..2019-05-26T13:29:13.367Z)
User #1 sees two events separated by 12 seconds. With standard sessions, the gap defaults to 10 seconds; both scores are in different sessions, so the scores aren't added.
User #2 sees four events, seperated by two, seven, and three seconds, respectively. Since none of the gaps are greater than the default, the four events are in the same standard session and added together (18 points).
The dynamic sessions specify a five-second gap, so they use the following windows and scores:
user=user-2, score=4, window=[2019-05-26T14:30:22.969Z..2019-05-26T14:30:32.969Z) user=user-1, score=9, window=[2019-05-26T14:30:22.429Z..2019-05-26T14:30:30.553Z) user=user-1, score=9, window=[2019-05-26T14:30:33.276Z..2019-05-26T14:30:41.849Z) user=user-2, score=10, window=[2019-05-26T14:30:37.357Z..2019-05-26T14:30:47.357Z)
With dynamic sessions, User #2 gets different scores. The third messages arrives seven seconds after the second message, so it's grouped into a different session. The large, 18-point session is split into two 9-point sessions.