Spark Stream new Job after stream start - java

I have a situation where I am trying to stream using spark streaming from kafka. The stream is a direct stream. I am able to create a stream and then start streaming, also able to get any updates (if any) on kafka via the streaming.
The issue comes in when i have a new request to stream a new topic. Since SparkStreaming context can be only 1 per jvm, I cannot create a new stream for every new request.
The way I figured out is
Once a DStream is created and spark streaming is already in progress, just attach a new stream to it. This does not seem to work, the createDStream (for a new topic2) does not return a stream and further processing is stopped. The streaming keep on continuing on the first request (say topic1).
Second, I thought to stop the stream, create DStream and then start streaming again. I cannot use the same streaming context (it throws an excpection that jobs cannot be added after streaming has been stopped), and if I create a new stream for new topic (topic2), the old stream topic (topic1) is lost and it streams only the new one.
Here is the code, have a look
JavaStreamingContext javaStreamingContext;
if(null == javaStreamingContext) {
javaStreamingContext = JavaStreamingContext(sparkContext, Durations.seconds(duration));
} else {
StreamingContextState streamingContextState = javaStreamingContext.getState();
if(streamingContextState == StreamingContextState.STOPPED) {
javaStreamingContext = JavaStreamingContext(sparkContext, Durations.seconds(duration));
}
}
Collection<String> topics = Arrays.asList(getTopicName(schemaName));
SparkVoidFunctionImpl impl = new SparkVoidFunctionImpl(getSparkSession());
KafkaUtils.createDirectStream(javaStreamingContext,
LocationStrategies.PreferConsistent(),
ConsumerStrategies.<String, String>Subscribe(topics, getKafkaParamMap()))
.map((stringStringConsumerRecord) -> stringStringConsumerRecord.value())
.foreachRDD(impl);
if (javaStreamingContext.getState() == StreamingContextState.ACTIVE) {
javaStreamingContext.start();
javaStreamingContext.awaitTermination();
}
Don't worry about SparkVoidFunctionImpl, this is a custom class with is the implementation of VoidFunction.
The above is approach 1, where i do not stop the existing streaming. When a new request comes into this method, it does not get a new streaming object, it tries to create a dstream. The issue is the DStream object is never returned.
KafkaUtils.createDirectStream(javaStreamingContext,
LocationStrategies.PreferConsistent(),
ConsumerStrategies.<String, String>Subscribe(topics, getKafkaParamMap()))
This does not return a dstream, the control just terminates without an error.The steps further are not executed.
I have tried many things and read multiple article, but I belive this is a very common production level issue. Any streaming done is to be done on multiple different topics and each of them is handled differently.
Please help

The thing is spark master sends out code to workers and although the data is streaming, underlying code and variable values remain static unless job is restarted.
Few options I could think:
Spark Job server: Every time you want to subscribe/stream from a different topic instead of touching already running job, start a new job. From your API body you can supply the parameters or topic name. If you want to stop streaming from a specific topic, just stop respective job. It will give you a lot of flexibility and control on resources.
[Theoritical] Topic Filter: Subscribe all topics you think you will want, when records are pulled for a duration, filter out records based on a LIST of topics. Manipulate this list of topics through API to increase or decrease your scope of topics, it could be a broadcast variable as well. This is just an idea, I have not tried this option at all.
Another work around is to relay your Topic-2 data to Topic-1 using a microservice whenever you need it & stop if you don't want to.

Related

Batch consumer camel kafka

I am unable to read in batch with the kafka camel consumer, despite following an example posted here. Are there changes I need to make to my producer, or is the problem most likely with my consumer configuration?
The application in question utilizes the kafka camel component to ingest messages from a rest endpoint, validate them, and place them on a topic. I then have a separate service that consumes them from the topic and persists them in a time-series database.
The messages were being produced and consumed one at a time, but the database expects the messages to be consumed and committed in batch for optimal performance. Without touching the producer, I tried adjusting the consumer to match the example in the answer to this question:
How to transactionally poll Kafka from Camel?
I wasn't sure how the messages would appear, so for now I'm just logging them:
from(kafkaReadingConsumerEndpoint).routeId("rawReadingsConsumer").process(exchange -> {
// simple approach to generating errors
String body = exchange.getIn().getBody(String.class);
if (body.startsWith("error")) {
throw new RuntimeException("can't handle the message");
}
log.info("BODY:{}", body);
}).process(kafkaOffsetManager);
But the messages still appear to be coming across one at a time with no batch read.
My consumer config is this:
kafka:
host: myhost
port: myport
consumer:
seekTo: beginning
maxPartitionFetchBytes: 55000
maxPollRecords: 50
consumerCount: 1
autoOffsetReset: earliest
autoCommitEnable: false
allowManualCommit: true
breakOnFirstError: true
Does my config need work, or are there changes I need to make to the producer to have this work correctly?
At the lowest layer, the KafkaConsumer#poll method is going to return an Iterator<ConsumerRecord>; there's no way around that.
I don't have in-depth experience with Camel, but in order to get a "batch" of records, you'll need some intermediate collection to "queue" the data that you want to eventually send downstream to some "collection consumer" process. Then you will need some "switch" processor that says "wait, process this batch" or "continue filling this batch".
As far as databases go, that process is exactly what Kafka Connect JDBC Sink does with batch.size config.
We solved a similar requirement by using the Aggregation [1] capability provided by Camel
A rough code snippet
#Override
public void configure() throws Exception {
// 1. Define your Aggregation Strat
AggregationStrategy agg = AggregationStrategies.flexible(String.class)
.accumulateInCollection(ArrayList.class)
.pick(body());
from("kafka:your-topic?and-other-params")
// 2. Define your Aggregation Strat Params
.aggregate(constant(true), agg)
.completionInterval(1000)
.completionSize(100)
.parallelProcessing(true)
// 3. Generate bulk insert statement
.process(exchange -> {
List<String> body = (List<String>) exchange.getIn().getBody();
String query = generateBulkInsertQueryStatement("target-table", body);
exchange.getMessage().setBody(query);
})
.to("jdbc:dataSource");
}
There are a variety of strategies that you can implement, but we chose this particular one because it allows you to create a List of strings for the message contents that we need to ingest into the db. [2]
We set a variety of different params such as completionInterval & completionSize. The most important one for us was to set parallellProcessing(true) [3] ; without that our performance wasn't nearly getting the required throughput.
Once the aggregation has either collected 100 messages or 1000 ms has passed, then the processor generates a bulk insert statement, which then gets sent to the db.
[1] https://camel.apache.org/components/3.18.x/eips/aggregate-eip.html
[2] https://camel.apache.org/components/3.18.x/eips/aggregate-eip.html#_aggregating_into_a_list
[3] https://camel.apache.org/components/3.18.x/eips/aggregate-eip.html#_worker_pools

passing an Akka stream to an upstream service to populate

I need to call an upstream service (Azure Blob Service) to push data to an OutputStream, which then i need to turn around and push it back to the client, thru akka. Without akka (and just servlet code), i'd just get the ServletOutputStream and pass it to the azure service's method.
The closest i can try to stumble upon, and clearly this is wrong, is something like this
Source<ByteString, OutputStream> source = StreamConverters.asOutputStream().mapMaterializedValue(os -> {
blobClient.download(os);
return os;
});
ResponseEntity resposeEntity = HttpEntities.create(ContentTypes.APPLICATION_OCTET_STREAM, preAuthData.getFileSize(), source);
sender().tell(new RequestResult(resposeEntity, StatusCodes.OK), self());
The idea is i'm calling an upstream service to get an outputstream populated by calling
blobClient.download(os);
It seems like the the lambda function gets called and returns, but then afterwards it fails, because there's no data or something. As if i'm not supposed to be have that lambda function do the work, but perhaps return some object that does the work? Not sure.
How does one do this?
The real issue here is that the Azure API is not designed for back-pressuring. There is no way for the output stream to signal back to Azure that it is not ready for more data. To put it another way: if Azure pushes data faster than you are able to consume it, there will have to be some ugly buffer overflow failure somewhere.
Accepting this fact, the next best thing we can do is:
Use Source.lazySource to only start downloading data when there is downstream demand (aka. the source is being run and data is being requested).
Put the download call in some other thread so that it continues executing without blocking the source from being returned. Once way to do this is with a Future (I'm not sure what Java best practices are, but should work fine either way). Although it won't matter initially, you may need to choose an execution context other than system.dispatcher - it all depends on whether download is blocking or not.
I apologize in advance if this Java code is malformed - I use Akka with Scala, so this is all from looking at the Akka Java API and Java syntax reference.
ResponseEntity responseEntity = HttpEntities.create(
ContentTypes.APPLICATION_OCTET_STREAM,
preAuthData.getFileSize(),
// Wait until there is downstream demand to intialize the source...
Source.lazySource(() -> {
// Pre-materialize the outputstream before the source starts running
Pair<OutputStream, Source<ByteString, NotUsed>> pair =
StreamConverters.asOutputStream().preMaterialize(system);
// Start writing into the download stream in a separate thread
Futures.future(() -> { blobClient.download(pair.first()); return pair.first(); }, system.getDispatcher());
// Return the source - it should start running since `lazySource` indicated demand
return pair.second();
})
);
sender().tell(new RequestResult(responseEntity, StatusCodes.OK), self());
The OutputStream in this case is the "materialized value" of the Source and it will only be created once the stream is run (or "materialized" into a running stream). Running it is out of your control since you hand the Source to Akka HTTP and that will later actually run your source.
.mapMaterializedValue(matval -> ...) is usually used to transform the materialized value but since it is invoked as a part of materialization you can use that to do side effects such as sending the matval in a message, just like you have figured out, there isn't necessarily anything wrong with that even if it looks funky. It is important to understand that the stream will not complete its materialization and become running until that lambda completes. This means problems if download() is blocking rather than forking off some work on a different thread and immediately returning.
There is however another solution: Source.preMaterialize(), it materializes the source and gives you a Pair of the materialized value and a new Source that can be used to consume the already started source:
Pair<OutputStream, Source<ByteString, NotUsed>> pair =
StreamConverters.asOutputStream().preMaterialize(system);
OutputStream os = pair.first();
Source<ByteString, NotUsed> source = pair.second();
Note that there are a few additional things to think of in your code, most importantly if the blobClient.download(os) call blocks until it is done and you call that from the actor, in that case you must make sure that your actor does not starve the dispatcher and stop other actors in your application from executing (see Akka docs: https://doc.akka.io/docs/akka/current/typed/dispatchers.html#blocking-needs-careful-management ).

Kafka Streams: Store is not ready

We recently upgraded Kafka to v1.1 and Confluent to v4.0.But upon upgrading we have encountered a persistent problems regarding state stores. Our application starts a collection of streams and we check for the state stores to be ready before killing the application after 100 tries. But after the upgrade there's atleast one stream that will have Store is not ready : the state store, <your stream>, may have migrated to another instance
The stream itself has RUNNING state and the messages will flow through but the state of the store still shows up as not ready. So I have no idea as to what may be happening.
Should I not check for store state?
And since our application has a lot of streams (~15), would starting
them simultaneously cause problems?
Should we not do a hard restart -- currently we run it as a service
on linux
We are running Kafka in cluster with 3 brokers.Below is a sample stream (not the entire code):
public BaseStream createStreamInstance() {
final Serializer<JsonNode> jsonSerializer = new JsonSerializer();
final Deserializer<JsonNode> jsonDeserializer = new JsonDeserializer();
final Serde<JsonNode> jsonSerde = Serdes.serdeFrom(jsonSerializer, jsonDeserializer);
MessagePayLoadParser<Note> noteParser = new MessagePayLoadParser<Note>(Note.class);
GenericJsonSerde<Note> noteSerde = new GenericJsonSerde<Note>(Note.class);
StreamsBuilder builder = new StreamsBuilder();
//below reducer will use sets to combine
//value1 in the reducer is what is already present in the store.
//value2 is the incoming message and for notes should have max 1 item in it's list (since its 1 attachment 1 tag per row, but multiple rows per note)
Reducer<Note> reducer = new Reducer<Note>() {
#Override
public Note apply(Note value1, Note value2) {
value1.merge(value2);
return value1;
}
};
KTable<Long, Note> noteTable = builder
.stream(this.subTopic, Consumed.with(jsonSerde, jsonSerde))
.map(noteParser::parse)
.groupByKey(Serialized.with(Serdes.Long(), noteSerde))
.reduce(reducer);
noteTable.toStream().to(this.pubTopic, Produced.with(Serdes.Long(), noteSerde));
this.stream = new KafkaStreams(builder.build(), this.properties);
return this;
}
There are some open questions here, like the ones Matthias put on comment, but will try to answer/give help to your actual questions:
Should I not check for store state?
Rebalancing is usually the case here. But in that case, you should not see that partition's thread keep consuming, but that processing should be "transferred" to be done to another thread that took over. Make sure if it is actually that very thread the one that keeps on processing that partition, and not the new one. Check kafka-consumer-groups utility to follow the consumers (threads) there.
And since our application has a lot of streams (~15), would starting them simultaneously cause problems? No, rebalancing is automatic.
Should we not do a hard restart -- currently we run it as a service on linux Are you keeping your state stores in a certain, non-default directory? You should configure your state stores directory properly and make sure it is accessible, insensitive to application restarts. Unsure about how you perform your hard restart, but some exception handling code should cover against it, closing your streams application.

kinesis getting data from multiple shards

I am trying to build a simple application that reads data from AWS Kinesis. I have managed to read data using a single shard but I want to get data from 4 different shards.
Problem is, I have a while loop which iterates as long as the shard is active which prevents me from reading data from different shards. So far I couldn't find an alternative algorithm nor was able to implement a KCL-based solution.
Many thanks in advance
public static void DoSomething() {
AmazonKinesisClient client = new AmazonKinesisClient();
//noinspection deprecation
client.setEndpoint(endpoint, serviceName, regionId);
/** get shards from the stream using describe stream method*/
DescribeStreamRequest describeStreamRequest = new DescribeStreamRequest();
describeStreamRequest.setStreamName(streamName);
List<Shard> shards = new ArrayList<>();
String exclusiveStartShardId = null;
do {
describeStreamRequest.setExclusiveStartShardId(exclusiveStartShardId);
DescribeStreamResult describeStreamResult = client.describeStream(describeStreamRequest);
shards.addAll(describeStreamResult.getStreamDescription().getShards());
if (describeStreamResult.getStreamDescription().getHasMoreShards() && shards.size() > 0) {
exclusiveStartShardId = shards.get(shards.size() - 1).getShardId();
} else {
exclusiveStartShardId = null;
}
}while (exclusiveStartShardId != null);
/** shards obtained */
String shardIterator;
GetShardIteratorRequest getShardIteratorRequest = new GetShardIteratorRequest();
getShardIteratorRequest.setStreamName(streamName);
getShardIteratorRequest.setShardId(shards.get(0).getShardId());
getShardIteratorRequest.setShardIteratorType("LATEST");
GetShardIteratorResult getShardIteratorResult = client.getShardIterator(getShardIteratorRequest);
shardIterator = getShardIteratorResult.getShardIterator();
GetRecordsRequest getRecordsRequest = new GetRecordsRequest();
while (!shardIterator.equals(null)) {
getRecordsRequest.setShardIterator(shardIterator);
getRecordsRequest.setLimit(250);
GetRecordsResult getRecordsResult = client.getRecords(getRecordsRequest);
List<Record> records = getRecordsResult.getRecords();
shardIterator = getRecordsResult.getNextShardIterator();
if(records.size()!=0) {
for(Record r : records) {
System.out.println(r.getPartitionKey());
}
}
try {
Thread.sleep(1000);
} catch (InterruptedException e) {
}
}
}
It is recommended that you will not read from a single process/worker from multiple shards. First, as you can see it is adding to the complexity of your code, but more importantly, you will have problems scaling up.
The "secret" of scalability is to have small and independent workers or other such units. Such design you can see in Hadoop, DynamoDB or Kinesis in AWS. It allows you to build small systems (micro-services), that can easily scale up and down as needed. You can easily add more units of work/data as your service becomes more successful, or other fluctuations in its usage.
As you can see in these AWS services, you sometimes can get this scalability automatically such in DynamoDB, and sometimes you need add shards to your kinesis streams. But for your application you need to control somehow your scalability.
In the case of Kinesis, you can scale up and down using AWS Lambda or Kinesis Client Library (KCL). Both of them are listening to the status of your streams (number of shards and events) and using it to add or remove workers and deliver the events for them to process. In both of these solutions you should build a worker that is working against a single shard.
If you need to align events from multiple shards, you can do that using some state service such as Redis or DynamoDB.
For a simpler and neater solution where you only have to worry about providing your own message processing code, I would recommend using the KCL Library.
Quoting from the documentation
The KCL acts as an intermediary between your record processing logic
and Kinesis Data Streams. The KCL performs the following tasks:
Connects to the data stream
Enumerates the shards within the data stream
Uses leases to coordinates shard associations with its workers
Instantiates a record processor for every shard it manages
Pulls data records from the data stream
Pushes the records to the corresponding record processor
Checkpoints processed records
Balances shard-worker associations (leases) when the worker instance count changes or when the data stream is resharded (shards are split or merged)

Spark Streaming get warn "replicated to only 0 peer(s) instead of 1 peers"

I use spark streaming to receive twitts from twitter.
I get many warning that says:
replicated to only 0 peer(s) instead of 1 peers
what is this warning for?
my code is:
SparkConf conf = new SparkConf().setAppName("Test");
JavaStreamingContext sc = new JavaStreamingContext(conf, Durations.seconds(5));
sc.checkpoint("/home/arman/Desktop/checkpoint");
ConfigurationBuilder cb = new ConfigurationBuilder();
cb.setOAuthConsumerKey("****************")
.setOAuthConsumerSecret("**************")
.setOAuthAccessToken("*********************")
.setOAuthAccessTokenSecret("***************");
JavaReceiverInputDStream<twitter4j.Status> statuses = TwitterUtils.createStream(sc,
AuthorizationFactory.getInstance(cb.build()));
JavaPairDStream<String, Long> hashtags = statuses.flatMapToPair(new GetHashtags());
JavaPairDStream<String, Long> hashtagsCount = hashtags.updateStateByKey(new UpdateReduce());
hashtagsCount.foreachRDD(new saveText(args[0], true));
sc.start();
sc.awaitTerminationOrTimeout(Long.parseLong(args[1]));
sc.stop();
When reading data with Spark Streaming, incoming data blocks are replicated to at least one another node/worker because of fault-tolerance. Without that it may happen that in case the runtime reads data from stream and then fails this particular piece of data would be lost (it's already read and erased from stream and it's also lost at the worker side because of failure).
Referring to the Spark documentation :
While a Spark Streaming driver program is running, the system receives
data from various sources and and divides it into batches. Each batch
of data is treated as an RDD, that is, an immutable parallel
collection of data. These input RDDs are saved in memory and
replicated to two nodes for fault-tolerance.
The warning in your case means that incoming data from stream are not replicated at all. The reason for that may be that you run the app with just one instance of Spark worker or running in local mode. Try to start more Spark workers and see if the warning is gone.

Categories