Kafka: Efficiently join windowed aggregates to events - apache-kafka-streams

I'm prototyping a fraud application. We'll frequently have metrics like "total amount of cash transactions in the last 5 days" that we need to compare against some threshold to determine if we raise an alert.
We're looking to use Kafka Streams to create and maintain the aggregates and then create an enhanced version of the incoming transaction that has the original transaction fields plus the aggregates. This enhanced record gets processed by a downstream rules system.
I'm wondering the best way to approach this. I've prototyped creating the aggregates with code like this:
TimeWindows twoDayHopping TimeWindows.of(TimeUnit.DAYS.toMillis(2))
.advanceBy(TimeUnit.DAYS.toMillis(1));
KStream<String, AdditiveStatistics> aggrStream = transactions
.filter((key,value)->{
return value.getAccountTypeDesc().equals("P") &&
value.getPrimaryMediumDesc().equals("CASH");
})
.groupByKey()
.aggregate(AdditiveStatistics::new,
(key,value,accumulator)-> {
return AdditiveStatsUtil
.advance(value.getCurrencyAmount(),accumulator),
twoDayHopping,
metricsSerde,
"sas10005_store")
}
.toStream()
.map((key,value)-> {
value.setTransDate(key.window().start());
return new KeyValue<String, AdditiveStatistics>(key.key(),value);
})
.through(Serdes.String(),metricsSerde,datedAggrTopic);;
This creates a store-backed stream that has a records per key per window. I then join the original transactions stream to this window to produce the final output to a topic:
JoinWindows joinWindow = JoinWindows.of(TimeUnit.DAYS.toMillis(1))
.before(TimeUnit.DAYS.toMillis(1))
.after(-1)
.until(TimeUnit.DAYS.toMillis(2)+1);
KStream<String,Transactions10KEnhanced> enhancedTrans = transactions.join(aggrStream,
(left,right)->{
Transactions10KEnhanced out = new Transactions10KEnhanced();
out.setAccountNumber(left.getAccountNumber());
out.setAccountTypeDesc(left.getAccountTypeDesc());
out.setPartyNumber(left.getPartyNumber());
out.setPrimaryMediumDesc(left.getPrimaryMediumDesc());
out.setSecondaryMediumDesc(left.getSecondaryMediumDesc());
out.setTransactionKey(left.getTransactionKey());
out.setCurrencyAmount(left.getCurrencyAmount());
out.setTransDate(left.getTransDate());
if(right != null) {
out.setSum2d(right.getSum());
}
return out;
},
joinWindow);
This produces the correct results, but it seems to run for quite a while, even with a low number of records. I'm wondering if there's a more efficient way to achieve the same result.

It's a config issues: cf http://docs.confluent.io/current/streams/developer-guide.html#memory-management
Disable caching by setting cache size to zero (parameter cache.max.bytes.buffering in StreamsConfig) will resolve the "delayed" delivery to the output topic.
You might also read this blog post for some background information about Streams design: https://www.confluent.io/blog/watermarks-tables-event-time-dataflow-model/

Related

Time semantics between KStream and KTable

I am trying to build the following topology:
Using Debezium Connectors, I am pulling 2 tables (let's called them tables A, and DA). As per DBZ, the topics where the table rows are stored have the structure { before: "...", after: "..." }.
First steps in my topology are to create "clean" KStreams off these two "table" topics. The sub-topology there looks roughly like this:
private static KStream<String, TABLE_A.Value> getTableARowByIdStream(
StreamsBuilder builder, Properties streamsConfig) {
return builder
.stream("TABLE_A", Consumed.withTimestampExtractor(Application::getRowDate))
.filter((key, envelope) -> [ some filtering condition ] )
.map((key, envelope) -> [ maps to TABLE_A.Value ] )
.through(tableRowByIdTopicName);
}
Notice that I am assigning the record time explicitly because the table rows will be CDC'ed "years" after they were originally published. What the function is doing at the moment is faking the time starting at 2010-01-01 and, using an AtomicInteger, adding 1 millisecond for each consumed entity. It does this for tables A but it doesn't for DA (I will explain why later).
Phase 2 of the topology is to build 1 KTable based on the "cleaned" topic for table A, like this:
private static KTable<String, EntityInfoList> getEntityInfoListById(
KStream<String, TABLE_A.Value> tableAByIdStream) {
return tableAByIdStream
.map((key, value) -> [ some mapping ] )
.groupByKey()
.aggregate(() -> [ builds up a EntityInfoList object ] ));
}
Finally, with th KTable ready, I'm joining them with the KStream over DA like so:
private static KStream<String, OutputTopicEntity> getOutputTopicEntityStream(
KStream<String, Table_DA.Value> tableDAStream,
KTable<String, EntityInfoList> tableA_KTable) {
KStream<String, Table_DA>[] branches = tableDAStream.branch(
(key, value) -> [ some logic ],
(key, value) -> true);
KStream<String, OutputTopicEntity> internalAccountRefStream = branches[0]
.join(
tableA_KTable,
(streamValue, tableValue) -> [ some logic to build a list of OutputTopicEntity ])
.flatMap((key, listValue) -> [ some logic to flatten it ]));
[ similar logic with branch[1] ]
}
My problem is, despite the fact that I am "faking" the time for records coming from the Table_A topic (I've verified that they are referencing 2010/01/01 using kafkacat) and entries in Table_DA (the stream side of the join) have timestamps around today '2019/08/14'), it doesn't seem like Kafka Streams is holding reading any of the entries from Table_DA KStream until it has ingested all records from Table_A into the KTable.
As a result of that, I don't have all the "join hits" that I was expecting and it is also nondeterministic. My understanding based on this sentence from What are the differences between KTable vs GlobalKTable and leftJoin() vs outerJoin()? was the opposite:
For stream-table join, Kafka Stream align record processing ordered based on record timestamps. Thus, the update to the table are aligned with the records of you stream.
My experience so far is this is not happening. I can also easily see how my application continues churning through the Table_A topic way after it has consumed all entries in Table_DA stream (it happens to be 10 times smaller).
Am I doing something wrong?
Timestamp synchronization is best effort before 2.1.0 release (cf. https://issues.apache.org/jira/browse/KAFKA-3514).
As of 2.1.0, timestamps are synchronized strictly. However, if one input does not have any data, Kafka Streams will "enforce" processing as described in KIP-353 to avoid blocking forever. If you have bursty inputs and want to "block" processing for some time if one input has no data, you can increase configuration parameter max.task.idle.ms (default is 0) as introduced in 2.1.0 via KIP-353.

Spring Kafka Stream doesn't get written

I'm writing a Spring Boot (2.1.4) app trying to use Spring Cloud Streams for Kafka.
What I'm trying to do is maintain a list of sensors on one topic ("sensors"). OTOH, I have incoming data on the other topic ("data"). What I'm trying to achieve is that when I get data for a sensor I don't already have, I want to add it to the sensor list.
To do that, I create a KTable<String, Sensor> from the sensors topic, map the temperature topic to the pure sensor's data (in this case, its name) and do an outer join with a ValueJoiner that retains the sensor if present, otherwise use the reading's sensor. Then, I write the result back to the sensors topic.
KTable<String, Sensor> sensorTable = ...;
KStream<String, SensorData> sensorDataStream = ...;
// get sensors providing measurements
KTable<String, Sensor> sensorsFromData =
sensorDataStream.groupByKey()
.aggregate(
Sensor::new,
(k, v, s) -> {
s.setName(k);
return s;
},
Materialized.with(Serdes.String(), SensorSerde.SERDE));
// join both sensor tables, preferring the existing ones
KTable<String, Sensor> joinedSensorTable =
sensorTable.outerJoin(
sensorsFromData,
// only use sensors from measurements if sensor not already present
(ex, ft) -> (ex != null) ? ex : ft,
Materialized.<String, Sensor, KeyValueStore<Bytes, byte[]>>as(SENSORS_TABLE)
.withKeySerde(Serdes.String()).withValueSerde(SensorSerde.SERDE));
// write to new topic for downstream services
joinedSensorTable.toStream();
This works fine if I create this using a StreamBuilder - i.e. if the sensorTable and sensorDataStream are coming from something like builder.table("sensors", Consumed.with(Serdes.String(), SensorSerde.SERDE)).
However, I'm trying to use Spring Stream Binding for this, ie the above code is wrapped in
#Configuration
#EnableBinding(SensorTableBinding.class)
class StreamConfiguration {
static final String SENSORS_TABLE = "sensors-table";
#StreamListener
#SendTo("sensorsOut")
private KStream<String, Sensor> getDataFromData
(#Input("sensors") KTable<String, Sensor> sensorTable,
#Input("data") KStream<String, SensorData> sensorDataStream) {
// ...
return joinedSensorTable.toStream();
}
}
with a
interface SensorTableBinding {
#Input("sensors")
KTable<String, Sensor> sensorStream();
#Output("sensorsOut")
KStream<String, Sensor> sensorOutput();
#Input("data")
KStream<String, SensorData> sensorDataStream();
}
Here is the spring stream section of the application.properties:
spring.cloud.stream.kafka.streams.binder.configuration.default.key.serde: org.apache.kafka.common.serialization.Serdes$StringSerde
spring.cloud.stream.kafka.streams.binder.configuration.default.value.serde: org.apache.kafka.common.serialization.Serdes$StringSerde
spring.cloud.stream.kafka.binder.brokers: ${spring.kafka.bootstrap-servers}
spring.cloud.stream.kafka.binder.configuration.auto.offset.reset: latest
spring.cloud.stream.kafka.binder.bindings.sensors.group: sensor-service
spring.cloud.stream.kafka.binder.bindings.sensors.destination: sensors
spring.cloud.stream.kafka.binder.bindings.sensorsOut.destination: sensors
spring.cloud.stream.kafka.binder.data.group: sensor-service
spring.cloud.stream.kafka.binder.data.destination: data
The stream gets initialized fine, and the join is performed (the key-value-store is filled properly), however, the resulting stream is never written to the "sensors" topic.
Why? Am I missing something?
Also: I'm sure there's a better way to de/serialize my objects from/to JSON using an existing Serde, rather than having to declare classes of my own to add to the processing (SensorSerde/SensorDataSerde are thin delegation wrapper to an ObjectMapper)?
Turns out the data was written after all, but to the wrong topic, namely sensorOut.
The reason was the configuration. Instead of
spring.cloud.stream.kafka.binder.bindings.sensors.destination: sensors
spring.cloud.stream.kafka.binder.bindings.sensorsOut.destination: sensors
the topics are configured with this:
spring.cloud.stream.bindings.sensors.destination: sensors
spring.cloud.stream.bindings.sensorsOut.destination: sensors
For the sensors and data topic, that didn't matter, because the binding's name was the same as the topic; but since Spring couldn't find a proper destination for the output, it used the binding's name sensorOut and wrote the data there.
As a note, the whole configuration setup around these is very confusing. The individual items are documented, but it's hard to tell for each to which configuration prefix they belong. Looking into the source code doesn't help either, because at that level what's passed around are Maps with the key stripped of the prefix at runtime, so it's really hard to tell where the data is coming from and what it will contain.
IMO it would really help to have acual #ConfigurationProperties-like data classes passed around, which would make it so much easier to understand.

Spring #StreamListener process(KStream<?,?> stream) Partition

I have a topic with multiple partitions in my stream processor i just wanted to stream that from one partition, and could nto figure out how to configure this
spring.cloud.stream.kafka.streams.bindings.input.consumer.application-id=s-processor
spring.cloud.stream.bindings.input.destination=uinput
spring.cloud.stream.bindings.input.group=r-processor
spring.cloud.stream.bindings.input.contentType=application/java-serialized-object
spring.cloud.stream.bindings.input.consumer.header-mode=raw
spring.cloud.stream.bindings.input.consumer.use-native-decoding=true
spring.cloud.stream.bindings.input.consumer.partitioned=true
#StreamListener(target = "input")
// #SendTo(value = { "uoutput" })
public void process(KStream<UUID, AModel> ustream) {
I want only one partition data to be processed by this processor, there will be other processors for other partition(s)
So far my finding is something to do with https://kafka.apache.org/20/javadoc/org/apache/kafka/streams/StreamsConfig.html#PARTITION_GROUPER_CLASS_CONFIG, but couldnot find how to set this property in spring application.properties
I think the partition grouper is to group partition with tasks within a single processor. If you want to ensure that only a single partition is processed by a processor, then you need to provide at least the same number of processor instances as the topic partitions. For e.g. if your topic has 4 partitions, then you need to have 4 instances of the stream application to ensure that each instance is only processing a single partition.
Kafka Streams does not allow to read a single partition. If you subscribe to a topic, all partitions are consumed and distributed over the available instances. Thus, you can't know in advance, which partition is assigned to what instance, and all instances execute the same code.
But each partition linked to processor has different kind of data hence require different processor application
For this case, the processor (or transformer) must be able to process data for all partitions. Kafka Streams exposes the partitions number via the ProcessorContext object that is handed to a processor via init() method: https://kafka.apache.org/20/javadoc/org/apache/kafka/streams/kstream/Transformer.html#init-org.apache.kafka.streams.processor.ProcessorContext-
Thus, you need to "branch" with within your transformer to apply different processing logic based on the partition:
ustream.transform(() -> new MyTransformer());
class MyTransformer implement Transformer {
// other methods omitted
R transform(K key, V value) {
switch(context.partition()) { // get context from `init()`
case 0:
// your processing logic
break;
case 1:
// your processing logic
break;
// ...
}
}

Support method assignTimestampsAndWatermarks in KeyedStream of Flink DataStream

we are using flink to process data from kafka topic, the data in kafka topic comes from massive IoT devices, we hash those devices into different partition based on hash code of device ID, so each partition has event data from many devices, and those event data contains "timestamp", "value", "deviceID" and so on. we want to use flink to do streaming process of those data, and our rule is to calculate the average value in 20 seconds for each device. but this is one problem, since all device share same timestamp and watermark generator, and each device has clock skew, the skew maybe very large, so this breaks rule of strictly ascending event time, if we can support assignTimestampsAndWatermarks in KeyedStream, then all those problem could be resolved.
Sample code:
DataStream<Order> timestampStream = orderStream.assignTimestampsAndWatermarks(new BoundedOutOfOrdernessGenerator());
KeyedStream<Order, String> keyStream = timestampStream.keyBy(new KeySelector<Order,String>() {
public String getKey(Order value) throws Exception {
return value.getId();
}
});
keyStream.window(TumblingEventTimeWindows.of(Time.seconds(20)));

Meteor: filter data in publish or on client

In Meteor I want to work on the document level when having a Mongo database and according to sources, what I have to watch out for is expensive publications so today my question is:
How would I go about publishing documents with relations, would I follow the relational-type of query where we would find assignment details with an assignment id like this:
Meteor.publish('someName', function () {
var empId = "dj4nfhd56k7bhb3b732fd73fb";
var assignmentData = Assignment.find({ employee_id: empId });
return AssignmentDetails.find({ assignment_id: $in [ assignment ] });
});
or should we rather take an approach like this where we skip the filtering step in the publish and instead publish every assignment_detail and handle that filter on the client:
Meteor.publish('someName', function () {
var empId = "dj4nfhd56k7bhb3b732fd73fb";
var assignmentData = Assignment.find({ employee_id: empId });
var detailData = AssignmentDetails.find({ employee_id: empId });
return [ assignmentData, detailData];
});
I guess this is a question of whether the amount of data being searched trough on the server should be more then or if the amount of data being transferred to the client should be bigger.
Which of these would be most cost effective for the server?
It's a matter of opinion, but if possible I would strongly recommend attaching employee_id to docs in AssignmentDetails, as you have in the second example. You're correct in suggesting that publications are expensive, but much more so if the publication function is more complex than necessary, and you can reduce your pub function to one line if you have employee_id in AssignmentDetails (even where there are many employee_ids for each assignment) by just searching on that. You don't even need to return that field to the client (you can specify the fields to return in your find), so the only incurred overhead would be in database storage (which is v. cheap) and adding it to inserted/updated AssignmentDetails docs (which would be imperceptible). The actual amount of data transferred would be the same as in the first case.
The alternative of just publishing everything might be fine for a small collection, but it really depends on the number of assignments, and it's not going to be at all scalable this way. You need to send the entire collection to the client every time a client connects, which is expensive and time-consuming at both ends if it's more than a MB or so, and there isn't really any way round that overhead when you're talking about a dynamic (i.e. frequently-changing) collection, which I think you are (whereas for largely static collections you can do things with localStorage and poll-and-diff).

Resources