Hi Jose,
You are right about using {{applyToKeyedState}} to register timers under
{{KeyedBroadcastProcessFunction}}. I think the author overlooked this, and
we could update the document.
As for specifying a smaller keyset to apply function, I think this is a
possible direction of optimization. This needs more consideration and
discussion in the community.
Best,
Zakelly
On Thu, Aug 29, 2024 at 11:55 AM Zakelly Lan <zakelly....@gmail.com> wrote:
> Hi Jose,
>
> You are right about using {{applyToKeyedState}} to register timers under
> {{KeyedBroadcastProcessFunction}}. I think the author overlooked this, and
> we could update the document.
>
> As for specifying a smaller keyset to apply function, I think this is a
> possible direction of optimization. This needs more consideration and
> discussion in the community.
>
>
> Best,
> Zakelly
>
> On Thu, Aug 29, 2024 at 4:27 AM Jose Vargas Badilla via user <
> user@flink.apache.org> wrote:
>
>> On a slight tangent, it seems to me that the DataStream v2 API only
>> exposes the ability to apply a function to all partitions, not a subset,
>> which the current applyToKeyedState does allow. Is that accurate? If so,
>> adding that ability to the v2 API would be helpful. For example, there can
>> be a process function with multiple keyed states whose keysets differ by
>> several orders of magnitude. In such cases, a user may want to apply a
>> function on just the smaller keyset.
>>
>> Thanks,
>> Jose
>>
>> On Wed, Aug 28, 2024 at 10:47 AM Jose Vargas Badilla <
>> jose.var...@datadoghq.com> wrote:
>>
>>> Hi,
>>>
>>> I recently learned that timers can be set in the KeyedStateFunction that
>>> is passed to KeyedBroadcastProcessFunction.Context#applyToKeyedState. The
>>> "trick" is to store a reference to the timerService that is available in
>>> processElement.
>>>
>>> This is behavior I have not seen explicitly documented before. The
>>> closest I could find is
>>>
>>> Context in the processBroadcastElement() method contains the method
>>> applyToKeyedState(StateDescriptor<S,
>>>> VS> stateDescriptor, KeyedStateFunction<KS, S> function). This allows
>>>> to register a KeyedStateFunction to be applied to all states of all
>>>> keys associated with the provided stateDescriptor.
>>>
>>>
>>> from
>>> https://nightlies.apache.org/flink/flink-docs-release-1.19/docs/dev/datastream/fault-tolerance/broadcast_state/#broadcastprocessfunction-and-keyedbroadcastprocessfunction
>>>
>>>
>>> A developer would need to know that timers are implied by *all states*.
>>> However, the section immediately after that says
>>>
>>> Registering timers is only possible at processElement() of the
>>>> KeyedBroadcastProcessFunction and only there. It is not possible in
>>>> the processBroadcastElement() method, as there is no key associated to
>>>> the broadcasted elements.
>>>
>>>
>>> which made me think that setting timers *anywhere* in
>>> processBroadcastElement, including in the user supplied KeyedStateFunction,
>>> is not possible, even though it is.
>>>
>>>
>>> *Should the documentation be updated to mention that timers can be set
>>> from applyToKeyedState?*
>>> I did notice that the new DataStream v2 API makes this behavior much
>>> clearer via the new ApplyPartitionFunction#apply(Collector<OUT> collector,
>>> PartitionedContext ctx) method, which is great!
>>>
>>> Below is an example of how to register timers in Flink 1.17.
>>>
>>> import lombok.NoArgsConstructor;
>>> import lombok.Setter;
>>> import org.apache.flink.api.common.state.State;
>>> import org.apache.flink.api.common.state.ValueState;
>>> import org.apache.flink.api.common.state.ValueStateDescriptor;
>>> import org.apache.flink.api.common.typeinfo.Types;
>>> import org.apache.flink.configuration.Configuration;
>>> import org.apache.flink.runtime.state.KeyedStateFunction;
>>> import org.apache.flink.streaming.api.TimerService;
>>> import
>>> org.apache.flink.streaming.api.functions.co.KeyedBroadcastProcessFunction;
>>> import org.apache.flink.util.Collector;
>>>
>>> /**
>>> * Function that shows that timers can be set in processBroadcastElement
>>> via ctx.applyToKeyedState.
>>> */
>>> public class ApplyToKeyedStateTimerExample
>>> extends KeyedBroadcastProcessFunction<String, String, Long, String> {
>>> private static final long serialVersionUID = -8682056773621520927L;
>>> private ValueState<String> state;
>>> private CachedTimerService cachedTimerService = new CachedTimerService();
>>>
>>> @Override
>>> public void processElement(String value, ReadOnlyContext ctx,
>>> Collector<String> out)
>>> throws Exception {
>>> state.update(value);
>>> cachedTimerService.setTimerService(ctx.timerService());
>>> }
>>>
>>> @Override
>>> public void processBroadcastElement(Long value, Context ctx,
>>> Collector<String> out)
>>> throws Exception {
>>> ctx.applyToKeyedState(
>>> new ValueStateDescriptor<>("state", Types.STRING),
>>> cachedTimerService.registerTimer(value));
>>> }
>>>
>>> @Override
>>> public void onTimer(final long timestamp, final OnTimerContext ctx, final
>>> Collector<String> out)
>>> throws Exception {
>>> var value = state.value();
>>> if (value != null) {
>>> out.collect(value);
>>> state.clear();
>>> }
>>> }
>>>
>>> @Override
>>> public void open(Configuration parameters) throws Exception {
>>> state = getRuntimeContext().getState(new
>>> ValueStateDescriptor<>("state", Types.STRING));
>>> }
>>>
>>> @NoArgsConstructor
>>> private static final class CachedTimerService {
>>> @Setter private TimerService timerService;
>>>
>>> public <K, S extends State> KeyedStateFunction<K, S> registerTimer(long
>>> timestamp) {
>>> return (key, state) -> {
>>> if (timerService != null) {
>>> timerService.registerEventTimeTimer(timestamp);
>>> }
>>> };
>>> }
>>> }
>>> }
>>>
>>> Here is a unit test for the above function.
>>>
>>> import java.util.stream.Collectors;
>>>
>>> import org.apache.flink.api.common.state.MapStateDescriptor;
>>> import org.apache.flink.api.common.typeinfo.Types;
>>> import org.apache.flink.streaming.runtime.streamrecord.StreamRecord;
>>> import org.apache.flink.streaming.util.ProcessFunctionTestHarnesses;
>>> import org.junit.jupiter.api.Assertions;
>>> import org.junit.jupiter.api.Test;
>>>
>>> public class ApplyToKeyedStateTimerExampleTest {
>>> @Test
>>> public void testTimersFromBroadcast() throws Exception {
>>> var harness =
>>> ProcessFunctionTestHarnesses.forKeyedBroadcastProcessFunction(
>>> new ApplyToKeyedStateTimerExample(),
>>> x -> x,
>>> Types.STRING,
>>> new MapStateDescriptor<>("foo", Types.STRING, Types.LONG));
>>> harness.open();
>>> harness.processElement("foo", 1);
>>> harness.processElement("bar", 1);
>>> harness.processBroadcastElement(2L, 2);
>>> harness.watermark(2);
>>> var output =
>>> harness.getOutput().stream()
>>> .filter(x -> x instanceof StreamRecord) // Filter out the
>>> watermark StreamElement
>>> .collect(Collectors.toList());
>>> Assertions.assertEquals(2, output.size());
>>> }
>>> }
>>>
>>>
>>> Thanks,
>>> Jose
>>>
>>
