Window (Google Cloud Dataflow SDK 1.9.1 API)

Google Cloud Dataflow SDK for Java, version 1.9.1

com.google.cloud.dataflow.sdk.transforms.windowing

Class Window


  • public class Window
extends Object
    Window logically divides up or groups the elements of a PCollection into finite windows according to a WindowFn. The output of Window contains the same elements as input, but they have been logically assigned to windows. The next GroupByKeys, including one within composite transforms, will group by the combination of keys and windows.

    See GroupByKey for more information about how grouping with windows works.

    Windowing

    Windowing a PCollection divides the elements into windows based on the associated event time for each element. This is especially useful for PCollections with unbounded size, since it allows operating on a sub-group of the elements placed into a related window. For PCollections with a bounded size (aka. conventional batch mode), by default, all data is implicitly in a single window, unless Window is applied.

    For example, a simple form of windowing divides up the data into fixed-width time intervals, using FixedWindows. The following example demonstrates how to use Window in a pipeline that counts the number of occurrences of strings each minute: 

     
 PCollection<String> items = ...;
 PCollection<String> windowed_items = items.apply(
   Window.<String>into(FixedWindows.of(Duration.standardMinutes(1))));
 PCollection<KV<String, Long>> windowed_counts = windowed_items.apply(
   Count.<String>perElement());

    Let (data, timestamp) denote a data element along with its timestamp. Then, if the input to this pipeline consists of {("foo", 15s), ("bar", 30s), ("foo", 45s), ("foo", 1m30s)}, the output will be {(KV("foo", 2), 1m), (KV("bar", 1), 1m), (KV("foo", 1), 2m)}

    Several predefined WindowFns are provided:

    • FixedWindows partitions the timestamps into fixed-width intervals.
    • SlidingWindows places data into overlapping fixed-width intervals.
    • Sessions groups data into sessions where each item in a window is separated from the next by no more than a specified gap.

    Additionally, custom WindowFns can be created, by creating new subclasses of WindowFn.

    Triggers

    Window.Bound.triggering(TriggerBuilder) allows specifying a trigger to control when (in processing time) results for the given window can be produced. If unspecified, the default behavior is to trigger first when the watermark passes the end of the window, and then trigger again every time there is late arriving data.

    Elements are added to the current window pane as they arrive. When the root trigger fires, output is produced based on the elements in the current pane.

    Depending on the trigger, this can be used both to output partial results early during the processing of the whole window, and to deal with late arriving in batches.

    Continuing the earlier example, if we wanted to emit the values that were available when the watermark passed the end of the window, and then output any late arriving elements once-per (actual hour) hour until we have finished processing the next 24-hours of data. (The use of watermark time to stop processing tends to be more robust if the data source is slow for a few days, etc.) 

     
 PCollection<String> items = ...;
 PCollection<String> windowed_items = items.apply(
   Window.<String>into(FixedWindows.of(Duration.standardMinutes(1)))
      .triggering(
          AfterWatermark.pastEndOfWindow()
              .withLateFirings(AfterProcessingTime
                  .pastFirstElementInPane().plusDelayOf(Duration.standardHours(1))))
      .withAllowedLateness(Duration.standardDays(1)));
 PCollection<KV<String, Long>> windowed_counts = windowed_items.apply(
   Count.<String>perElement());

    On the other hand, if we wanted to get early results every minute of processing time (for which there were new elements in the given window) we could do the following: 

     
 PCollection<String> windowed_items = items.apply(
   Window.<String>into(FixedWindows.of(Duration.standardMinutes(1))
      .triggering(
      .triggering(
          AfterWatermark.pastEndOfWindow()
              .withEarlyFirings(AfterProcessingTime
                  .pastFirstElementInPane().plusDelayOf(Duration.standardMinutes(1))))
      .withAllowedLateness(Duration.ZERO));

    After a GroupByKey the trigger is set to a trigger that will preserve the intent of the upstream trigger. See Trigger.getContinuationTrigger() for more information.

    See Trigger for details on the available triggers.

    • Constructor Detail

      • Window

        public Window()

    • Method Detail

      • into

        public static <T> Window.Bound<T> into(WindowFn<? super T,?> fn)
        Creates a Window PTransform that uses the given WindowFn to window the data.

        The resulting PTransform's types have been bound, with both the input and output being a PCollection<T>, inferred from the types of the argument WindowFn. It is ready to be applied, or further properties can be set on it first.

      • discardingFiredPanes

        @Experimental(value=TRIGGER)
public static <T> Window.Bound<T> discardingFiredPanes()
        Returns a new Window PTransform that uses the registered WindowFn and Triggering behavior, and that discards elements in a pane after they are triggered.

        Does not modify this transform. The resulting PTransform is sufficiently specified to be applied, but more properties can still be specified.

      • accumulatingFiredPanes

        @Experimental(value=TRIGGER)
public static <T> Window.Bound<T> accumulatingFiredPanes()
        Returns a new Window PTransform that uses the registered WindowFn and Triggering behavior, and that accumulates elements in a pane after they are triggered.

        Does not modify this transform. The resulting PTransform is sufficiently specified to be applied, but more properties can still be specified.

      • withAllowedLateness

        @Experimental(value=TRIGGER)
public static <T> Window.Bound<T> withAllowedLateness(Duration allowedLateness)
        Override the amount of lateness allowed for data elements in the pipeline. Like the other properties on this Window operation, this will be applied at the next GroupByKey. Any elements that are later than this as decided by the system-maintained watermark will be dropped.

        This value also determines how long state will be kept around for old windows. Once no elements will be added to a window (because this duration has passed) any state associated with the window will be cleaned up.

      • remerge

        public static <T> Window.Remerge<T> remerge()
        Creates a Window PTransform that does not change assigned windows, but will cause windows to be merged again as part of the next GroupByKey.


Oliko tästä sivusta apua? Kerro mielipiteesi

Palautteen aihe:

Tämä sivu
Dokumentaatio
Cloud Dataflow
Tuote
Tarvitsetko apua? Siirry tukisivullemme.