Process Function

The ProcessFunction

The ProcessFunction is a low-level stream processing operation, giving access to the basic building blocks of all (acyclic) streaming applications:

  • events (stream elements)
  • state (fault-tolerant, consistent, only on keyed stream)
  • timers (event time and processing time, only on keyed stream)

The ProcessFunction can be thought of as a FlatMapFunction with access to keyed state and timers. It handles events by being invoked for each event received in the input stream(s).

For fault-tolerant state, the ProcessFunction gives access to Flink’s keyed state, accessible via the RuntimeContext, similar to the way other stateful functions can access keyed state.

The timers allow applications to react to changes in processing time and in event time. Every call to the function processElement(...) gets a Context object which gives access to the element’s event time timestamp, and to the TimerService. The TimerService can be used to register callbacks for future event-/processing-time instants. With event-time timers, the onTimer(...) method is called when the current watermark is advanced up to or beyond the timestamp of the timer, while with processing-time timers, onTimer(...) is called when wall clock time reaches the specified time. During that call, all states are again scoped to the key with which the timer was created, allowing timers to manipulate keyed state.

If you want to access keyed state and timers you have to apply the ProcessFunction on a keyed stream:

  1. stream.keyBy(...).process(new MyProcessFunction());

Low-level Joins

To realize low-level operations on two inputs, applications can use CoProcessFunction or KeyedCoProcessFunction. This function is bound to two different inputs and gets individual calls to processElement1(...) and processElement2(...) for records from the two different inputs.

Implementing a low level join typically follows this pattern:

  • Create a state object for one input (or both)
  • Update the state upon receiving elements from its input
  • Upon receiving elements from the other input, probe the state and produce the joined result

For example, you might be joining customer data to financial trades, while keeping state for the customer data. If you care about having complete and deterministic joins in the face of out-of-order events, you can use a timer to evaluate and emit the join for a trade when the watermark for the customer data stream has passed the time of that trade.

Example

In the following example a KeyedProcessFunction maintains counts per key, and emits a key/count pair whenever a minute passes (in event time) without an update for that key:

  • The count, key, and last-modification-timestamp are stored in a ValueState, which is implicitly scoped by key.
  • For each record, the KeyedProcessFunction increments the counter and sets the last-modification timestamp
  • The function also schedules a callback one minute into the future (in event time)
  • Upon each callback, it checks the callback’s event time timestamp against the last-modification time of the stored count and emits the key/count if they match (i.e., no further update occurred during that minute)

This simple example could have been implemented with session windows. We use KeyedProcessFunction here to illustrate the basic pattern it provides.

Java

  1. import org.apache.flink.api.common.state.ValueState;
  2. import org.apache.flink.api.common.state.ValueStateDescriptor;
  3. import org.apache.flink.api.java.tuple.Tuple;
  4. import org.apache.flink.api.java.tuple.Tuple2;
  5. import org.apache.flink.configuration.Configuration;
  6. import org.apache.flink.streaming.api.functions.KeyedProcessFunction;
  7. import org.apache.flink.streaming.api.functions.KeyedProcessFunction.Context;
  8. import org.apache.flink.streaming.api.functions.KeyedProcessFunction.OnTimerContext;
  9. import org.apache.flink.util.Collector;
  12. // the source data stream
  13. DataStream<Tuple2<String, String>> stream = ...;
  15. // apply the process function onto a keyed stream
  16. DataStream<Tuple2<String, Long>> result = stream
  17.     .keyBy(value -> value.f0)
  18.     .process(new CountWithTimeoutFunction());
  20. /**
  21.  * The data type stored in the state
  22.  */
  23. public class CountWithTimestamp {
  25.     public String key;
  26.     public long count;
  27.     public long lastModified;
  28. }
  30. /**
  31.  * The implementation of the ProcessFunction that maintains the count and timeouts
  32.  */
  33. public class CountWithTimeoutFunction 
  34.         extends KeyedProcessFunction<Tuple, Tuple2<String, String>, Tuple2<String, Long>> {
  36.     /** The state that is maintained by this process function */
  37.     private ValueState<CountWithTimestamp> state;
  39.     @Override
  40.     public void open(OpenContext openContext) throws Exception {
  41.         state = getRuntimeContext().getState(new ValueStateDescriptor<>("myState", CountWithTimestamp.class));
  42.     }
  44.     @Override
  45.     public void processElement(
  46.             Tuple2<String, String> value, 
  47.             Context ctx, 
  48.             Collector<Tuple2<String, Long>> out) throws Exception {
  50.         // retrieve the current count
  51.         CountWithTimestamp current = state.value();
  52.         if (current == null) {
  53.             current = new CountWithTimestamp();
  54.             current.key = value.f0;
  55.         }
  57.         // update the state's count
  58.         current.count++;
  60.         // set the state's timestamp to the record's assigned event time timestamp
  61.         current.lastModified = ctx.timestamp();
  63.         // write the state back
  64.         state.update(current);
  66.         // schedule the next timer 60 seconds from the current event time
  67.         ctx.timerService().registerEventTimeTimer(current.lastModified + 60000);
  68.     }
  70.     @Override
  71.     public void onTimer(
  72.             long timestamp, 
  73.             OnTimerContext ctx, 
  74.             Collector<Tuple2<String, Long>> out) throws Exception {
  76.         // get the state for the key that scheduled the timer
  77.         CountWithTimestamp result = state.value();
  79.         // check if this is an outdated timer or the latest timer
  80.         if (timestamp == result.lastModified + 60000) {
  81.             // emit the state on timeout
  82.             out.collect(new Tuple2<String, Long>(result.key, result.count));
  83.         }
  84.     }
  85. }

Scala

  1. import org.apache.flink.api.common.state.ValueState
  2. import org.apache.flink.api.common.state.ValueStateDescriptor
  3. import org.apache.flink.api.java.tuple.Tuple
  4. import org.apache.flink.streaming.api.functions.KeyedProcessFunction
  5. import org.apache.flink.util.Collector
  7. // the source data stream
  8. val stream: DataStream[Tuple2[String, String]] = ...
  10. // apply the process function onto a keyed stream
  11. val result: DataStream[Tuple2[String, Long]] = stream
  12.   .keyBy(_._1)
  13.   .process(new CountWithTimeoutFunction())
  15. /**
  16.   * The data type stored in the state
  17.   */
  18. case class CountWithTimestamp(key: String, count: Long, lastModified: Long)
  20. /**
  21.   * The implementation of the ProcessFunction that maintains the count and timeouts
  22.   */
  23. class CountWithTimeoutFunction extends KeyedProcessFunction[Tuple, (String, String), (String, Long)] {
  25.   /** The state that is maintained by this process function */
  26.   lazy val state: ValueState[CountWithTimestamp] = getRuntimeContext
  27.     .getState(new ValueStateDescriptor[CountWithTimestamp]("myState", classOf[CountWithTimestamp]))
  30.   override def processElement(
  31.       value: (String, String), 
  32.       ctx: KeyedProcessFunction[Tuple, (String, String), (String, Long)]#Context, 
  33.       out: Collector[(String, Long)]): Unit = {
  35.     // initialize or retrieve/update the state
  36.     val current: CountWithTimestamp = state.value match {
  37.       case null =>
  38.         CountWithTimestamp(value._1, 1, ctx.timestamp)
  39.       case CountWithTimestamp(key, count, lastModified) =>
  40.         CountWithTimestamp(key, count + 1, ctx.timestamp)
  41.     }
  43.     // write the state back
  44.     state.update(current)
  46.     // schedule the next timer 60 seconds from the current event time
  47.     ctx.timerService.registerEventTimeTimer(current.lastModified + 60000)
  48.   }
  50.   override def onTimer(
  51.       timestamp: Long, 
  52.       ctx: KeyedProcessFunction[Tuple, (String, String), (String, Long)]#OnTimerContext, 
  53.       out: Collector[(String, Long)]): Unit = {
  55.     state.value match {
  56.       case CountWithTimestamp(key, count, lastModified) if (timestamp == lastModified + 60000) =>
  57.         out.collect((key, count))
  58.       case _ =>
  59.     }
  60.   }
  61. }

Python

  1. import datetime
  3. from pyflink.common import Row, WatermarkStrategy
  4. from pyflink.common.typeinfo import Types
  5. from pyflink.common.watermark_strategy import TimestampAssigner
  6. from pyflink.datastream import StreamExecutionEnvironment
  7. from pyflink.datastream.functions import KeyedProcessFunction, RuntimeContext
  8. from pyflink.datastream.state import ValueStateDescriptor
  9. from pyflink.table import StreamTableEnvironment
  12. class CountWithTimeoutFunction(KeyedProcessFunction):
  14.     def __init__(self):
  15.         self.state = None
  17.     def open(self, runtime_context: RuntimeContext):
  18.         self.state = runtime_context.get_state(ValueStateDescriptor(
  19.             "my_state", Types.PICKLED_BYTE_ARRAY()))
  21.     def process_element(self, value, ctx: 'KeyedProcessFunction.Context'):
  22.         # retrieve the current count
  23.         current = self.state.value()
  24.         if current is None:
  25.             current = Row(value.f1, 0, 0)
  27.         # update the state's count
  28.         current[1] += 1
  30.         # set the state's timestamp to the record's assigned event time timestamp
  31.         current[2] = ctx.timestamp()
  33.         # write the state back
  34.         self.state.update(current)
  36.         # schedule the next timer 60 seconds from the current event time
  37.         ctx.timer_service().register_event_time_timer(current[2] + 60000)
  39.     def on_timer(self, timestamp: int, ctx: 'KeyedProcessFunction.OnTimerContext'):
  40.         # get the state for the key that scheduled the timer
  41.         result = self.state.value()
  43.         # check if this is an outdated timer or the latest timer
  44.         if timestamp == result[2] + 60000:
  45.             # emit the state on timeout
  46.             yield result[0], result[1]
  49. class MyTimestampAssigner(TimestampAssigner):
  51.     def __init__(self):
  52.         self.epoch = datetime.datetime.utcfromtimestamp(0)
  54.     def extract_timestamp(self, value, record_timestamp) -> int:
  55.         return int((value[0] - self.epoch).total_seconds() * 1000)
  58. if __name__ == '__main__':
  59.     env = StreamExecutionEnvironment.get_execution_environment()
  60.     t_env = StreamTableEnvironment.create(stream_execution_environment=env)
  62.     t_env.execute_sql("""
  63.             CREATE TABLE my_source (
  64.               a TIMESTAMP(3),
  65.               b VARCHAR,
  66.               c VARCHAR
  67.             ) WITH (
  68.               'connector' = 'datagen',
  69.               'rows-per-second' = '10'
  70.             )
  71.         """)
  73.     stream = t_env.to_append_stream(
  74.         t_env.from_path('my_source'),
  75.         Types.ROW([Types.SQL_TIMESTAMP(), Types.STRING(), Types.STRING()]))
  76.     watermarked_stream = stream.assign_timestamps_and_watermarks(
  77.         WatermarkStrategy.for_monotonous_timestamps()
  78.                          .with_timestamp_assigner(MyTimestampAssigner()))
  80.     # apply the process function onto a keyed stream
  81.     result = watermarked_stream.key_by(lambda value: value[1]) \
  82.                                .process(CountWithTimeoutFunction()) \
  83.                                .print()
  84.     env.execute()

Before Flink 1.4.0, when called from a processing-time timer, the ProcessFunction.onTimer() method sets the current processing time as event-time timestamp. This behavior is very subtle and might not be noticed by users. Well, it’s harmful because processing-time timestamps are indeterministic and not aligned with watermarks. Besides, user-implemented logic depends on this wrong timestamp highly likely is unintendedly faulty. So we’ve decided to fix it. Upon upgrading to 1.4.0, Flink jobs that are using this incorrect event-time timestamp will fail, and users should adapt their jobs to the correct logic.

The KeyedProcessFunction

KeyedProcessFunction, as an extension of ProcessFunction, gives access to the key of timers in its onTimer(...) method.

Java

  1. @Override
  2. public void onTimer(long timestamp, OnTimerContext ctx, Collector<OUT> out) throws Exception {
  3.     K key = ctx.getCurrentKey();
  4.     // ...
  5. }

Scala

  1. override def onTimer(timestamp: Long, ctx: OnTimerContext, out: Collector[OUT]): Unit = {
  2.   var key = ctx.getCurrentKey
  3.   // ...
  4. }

Python

  1. def on_timer(self, timestamp: int, ctx: 'KeyedProcessFunction.OnTimerContext'):
  2.     key = ctx.get_current_key()
  3.     # ...

Timers

Both types of timers (processing-time and event-time) are internally maintained by the TimerService and enqueued for execution.

The TimerService deduplicates timers per key and timestamp, i.e., there is at most one timer per key and timestamp. If multiple timers are registered for the same timestamp, the onTimer() method will be called just once.

Flink synchronizes invocations of onTimer() and processElement(). Hence, users do not have to worry about concurrent modification of state.

Fault Tolerance

Timers are fault tolerant and checkpointed along with the state of the application. In case of a failure recovery or when starting an application from a savepoint, the timers are restored.

Checkpointed processing-time timers that were supposed to fire before their restoration, will fire immediately. This might happen when an application recovers from a failure or when it is started from a savepoint.

Timers are always asynchronously checkpointed, except for the combination of RocksDB backend / with incremental snapshots / with heap-based timers (will be resolved with FLINK-10026). Notice that large numbers of timers can increase the checkpointing time because timers are part of the checkpointed state. See the “Timer Coalescing” section for advice on how to reduce the number of timers.

Timer Coalescing

Since Flink maintains only one timer per key and timestamp, you can reduce the number of timers by reducing the timer resolution to coalesce them.

For a timer resolution of 1 second (event or processing time), you can round down the target time to full seconds. Timers will fire at most 1 second earlier but not later than requested with millisecond accuracy. As a result, there are at most one timer per key and second.

Java

  1. long coalescedTime = ((ctx.timestamp() + timeout) / 1000) * 1000;
  2. ctx.timerService().registerProcessingTimeTimer(coalescedTime);

Scala

  1. val coalescedTime = ((ctx.timestamp + timeout) / 1000) * 1000
  2. ctx.timerService.registerProcessingTimeTimer(coalescedTime)

Python

  1. coalesced_time = ((ctx.timestamp() + timeout) // 1000) * 1000
  2. ctx.timer_service().register_processing_time_timer(coalesced_time)

Since event-time timers only fire with watermarks coming in, you may also schedule and coalesce these timers with the next watermark by using the current one:

Java

  1. long coalescedTime = ctx.timerService().currentWatermark() + 1;
  2. ctx.timerService().registerEventTimeTimer(coalescedTime);

Scala

  1. val coalescedTime = ctx.timerService.currentWatermark + 1
  2. ctx.timerService.registerEventTimeTimer(coalescedTime)

Python

  1. coalesced_time = ctx.timer_service().current_watermark() + 1
  2. ctx.timer_service().register_event_time_timer(coalesced_time)

Timers can also be stopped and removed as follows:

Stopping a processing-time timer:

Java

  1. long timestampOfTimerToStop = ...;
  2. ctx.timerService().deleteProcessingTimeTimer(timestampOfTimerToStop);

Scala

  1. val timestampOfTimerToStop = ...
  2. ctx.timerService.deleteProcessingTimeTimer(timestampOfTimerToStop)

Python

  1. timestamp_of_timer_to_stop = ...
  2. ctx.timer_service().delete_processing_time_timer(timestamp_of_timer_to_stop)

Stopping an event-time timer:

Java

  1. long timestampOfTimerToStop = ...;
  2. ctx.timerService().deleteEventTimeTimer(timestampOfTimerToStop);

Scala

  1. val timestampOfTimerToStop = ...
  2. ctx.timerService.deleteEventTimeTimer(timestampOfTimerToStop)

Python

  1. timestamp_of_timer_to_stop = ...
  2. ctx.timer_service().delete_event_time_timer(timestamp_of_timer_to_stop)

Stopping a timer has no effect if no such timer with the given timestamp is registered.