Develop Reference

Retrieve from kafka last offsets for each partition in specified topic

Kafka gives useful command line tool kafka.tools.GetOffsetShell, but I need its functionality in my application.
I want to get all offsets for each partition in specified topic, like that:
bin/kafka-run-class.sh kafka.tools.GetOffsetShell --broker-list kafka:9092 --topic com.group.test.Foo
com.group.test.Foo:0:10
com.group.test.Foo:1:11
com.group.test.Foo:2:10
But I don't want to run process bin/kafka-run-class.sh kafka.tools.GetOffsetShell.
How can I do the same using kafka api in Java?
Do I have to create consumer and invoke: KafkaConsumer#position for each TopicPartition? I need simpler way?

By default, GetOffsetShell returns the end offset for each partitions. You could retrieve those offsets programmatically like this:
......
try (final KafkaConsumer<String, String> consumer = new KafkaConsumer<>(consumerProperties)) {
consumer.subscribe(Arrays.asList("topicName"));
Set<TopicPartition> assignment;
while ((assignment = consumer.assignment()).isEmpty()) {
consumer.poll(Duration.ofMillis(100));
}
consumer.endOffsets(assignment).forEach((tp, offset) -> System.out.println(tp + ": " + offset));
}

How to delete data which already been consumed by consumer? Kafka

I am doing data replication in kafka. But, the size of kafka log file is increases very quickly. The size reaches 5 gb in a day. As a solution of this problem, ı want to delete processed data immediately. I am using delete record method in AdminClient to delete offset. But when I look at the log file, data corresponding to that offset is not deleted.
RecordsToDelete recordsToDelete = RedcordsToDelete.beforeOffset(offset);
TopicPartition topicPartition = new TopicPartition(topicName,partition);
Map<TopicPartition,RecordsToDelete> deleteConf = new HashMap<>();
deleteConf.put(topicPartition,recordsToDelete);
adminClient.deleteRecords(deleteConf);
I don't want suggestions like (log.retention.hours , log.retention.bytes , log.segment.bytes , log.cleanup.policy=delete)
Because I just want to delete data consumed by the consumer. In this solution, I also deleted the data that is not consumed.
What are your suggestions?

You didn't do anything wrong. The code you provided works and I've tested it. Just in case I've overlooked something in your code, mine is:
public void deleteMessages(String topicName, int partitionIndex, int beforeIndex) {
TopicPartition topicPartition = new TopicPartition(topicName, partitionIndex);
Map<TopicPartition, RecordsToDelete> deleteMap = new HashMap<>();
deleteMap.put(topicPartition, RecordsToDelete.beforeOffset(beforeIndex));
kafkaAdminClient.deleteRecords(deleteMap);
}
I've used group: 'org.apache.kafka', name: 'kafka-clients', version: '2.0.0'
So check if you are targeting right partition ( 0 for the first one)
Check your broker version: https://kafka.apache.org/20/javadoc/index.html?org/apache/kafka/clients/admin/AdminClient.html says:
This operation is supported by brokers with version 0.11.0.0
Produce the messages from the same application, to be sure you're connected properly.
There is one more option you can consider. Using cleanup.policy=compact If your message keys are repeating you could benefit from it. Not just because older messages for that key will be automatically deleted but you can use the fact that message with null payload deletes all the messages for that key. Just don't forget to set delete.retention.ms and min.compaction.lag.ms to values small enough. In that case you can consume a message and than produce null payload for the same key ( but be cautious with this approach since this way you can delete messages ( with that key) you didn't consume)

Try this
DeleteRecordsResult result = adminClient.deleteRecords(recordsToDelete);
Map<TopicPartition, KafkaFuture<DeletedRecords>> lowWatermarks = result.lowWatermarks();
try {
for (Map.Entry<TopicPartition, KafkaFuture<DeletedRecords>> entry : lowWatermarks.entrySet()) {
System.out.println(entry.getKey().topic() + " " + entry.getKey().partition() + " " + entry.getValue().get().lowWatermark());
}
} catch (InterruptedException | ExecutionException e) {
e.printStackTrace();
}
adminClient.close();
In this code, you need to call entry.getValue().get().lowWatermark(), because adminClient.deleteRecords(recordsToDelete) returns a map of Futures, you need to wait for the Future to run by calling get()

This code will only work if the cleanup policy is "delete" or "compact, delete" else the code will throw a Policy Violation exception.

Apache Kafka System Error Handling

We are trying to implement Kafka as our message broker solution. We are deploying our Spring Boot microservices in IBM BLuemix, whose internal message broker implementation is Kafka version 0.10. Since my experience is more on the JMS, ActiveMQ end, I was wondering what should be the ideal way to handle system level errors in the java consumers?
Here is how we have implemented it currently
Consumer properties
enable.auto.commit=false
auto.offset.reset=latest
We are using the default properties for
max.partition.fetch.bytes
session.timeout.ms
Kafka Consumer
We are spinning up 3 threads per topic all having the same groupId, i.e one KafkaConsumer instance per thread. We have only one partition as of now. The consumer code looks like this in the constructor of the thread class
kafkaConsumer = new KafkaConsumer<String, String>(properties);
final List<String> topicList = new ArrayList<String>();
topicList.add(properties.getTopic());
kafkaConsumer.subscribe(topicList, new ConsumerRebalanceListener() {
#Override
public void onPartitionsRevoked(final Collection<TopicPartition> partitions) {
}
#Override
public void onPartitionsAssigned(final Collection<TopicPartition> partitions) {
try {
logger.info("Partitions assigned, consumer seeking to end.");
for (final TopicPartition partition : partitions) {
final long position = kafkaConsumer.position(partition);
logger.info("current Position: " + position);
logger.info("Seeking to end...");
kafkaConsumer.seekToEnd(Arrays.asList(partition));
logger.info("Seek from the current position: " + kafkaConsumer.position(partition));
kafkaConsumer.seek(partition, position);
}
logger.info("Consumer can now begin consuming messages.");
} catch (final Exception e) {
logger.error("Consumer can now begin consuming messages.");
}
}
});
The actual reading happens in the run method of the thread
try {
// Poll on the Kafka consumer every second.
final ConsumerRecords<String, String> records = kafkaConsumer.poll(1000);
// Iterate through all the messages received and print their
// content.
for (final TopicPartition partition : records.partitions()) {
final List<ConsumerRecord<String, String>> partitionRecords = records.records(partition);
logger.info("consumer is alive and is processing "+ partitionRecords.size() +" records");
for (final ConsumerRecord<String, String> record : partitionRecords) {
logger.info("processing topic "+ record.topic()+" for key "+record.key()+" on offset "+ record.offset());
final Class<? extends Event> resourceClass = eventProcessors.getResourceClass();
final Object obj = converter.convertToObject(record.value(), resourceClass);
if (obj != null) {
logger.info("Event: " + obj + " acquired by " + Thread.currentThread().getName());
final CommsEvent event = resourceClass.cast(converter.convertToObject(record.value(), resourceClass));
final MessageResults results = eventProcessors.processEvent(event
);
if ("Success".equals(results.getStatus())) {
// commit the processed message which changes
// the offset
kafkaConsumer.commitSync();
logger.info("Message processed sucessfully");
} else {
kafkaConsumer.seek(new TopicPartition(record.topic(), record.partition()), record.offset());
logger.error("Error processing message : {} with error : {},resetting offset to {} ", obj,results.getError().getMessage(),record.offset());
break;
}
}
}
}
// TODO add return
} catch (final Exception e) {
logger.error("Consumer has failed with exception: " + e, e);
shutdown();
}
You will notice the EventProcessor which is a service class which processes each record, in most cases commits the record in database. If the processor throws an error (System Exception or ValidationException) we do not commit but programatically set the seek to that offset, so that subsequent poll will return from that offset for that group id.
The doubt now is that, is this the right approach? If we get an error and we set the offset then until that is fixed no other message is processed. This might work for system errors like not able to connect to DB, but if the problem is only with that event and not others to process this one record we wont be able to process any other record. We thought of the concept of ErrorTopic where when we get an error the consumer will publish that event to the ErrorTopic and in the meantime it will keep on processing other subsequent events. But it looks like we are trying to bring in the design concepts of JMS (due to my previous experience) into kafka and there may be better way to solve error handling in kafka. Also reprocessing it from error topic may change the sequence of messages which we don't want for some scenarios
Please let me know how anyone has handled this scenario in their projects following the Kafka standards.
-Tatha

if the problem is only with that event and not others to process this one record we wont be able to process any other record
that's correct and your suggestion to use an error topic seems a possible one.
I also noticed that with your handling of onPartitionsAssigned you essentially do not use the consumer committed offset, as you seem you'll always seek to the end.
If you want to restart from the last succesfully committed offset, you should not perform a seek
Finally, I'd like to point out, though it looks like you know that, having 3 consumers in the same group subscribed to a single partition - means that 2 out of 3 will be idle.
HTH
Edo

Dataframes are slow to parse through small amount of data

I have 2 classes doing a similar task in Apache Spark but the one using data frame is many times slower than the "regular" one using RDD. (30x)
I would like to use data frame since it will eliminate a lot of code and classes we have but obviously I can't have it be that much slower.
The data set is nothing big. We have 30 some files with json data in each about events triggered from activities in another piece of software. There are between 0 to 100 events in each file.
A data set with 82 events will take about 5 minutes to be processed with data frames.
Sample code:
public static void main(String[] args) throws ParseException, IOException {
SparkConf sc = new SparkConf().setAppName("POC");
JavaSparkContext jsc = new JavaSparkContext(sc);
SQLContext sqlContext = new SQLContext(jsc);
conf = new ConfImpl();
HashSet<String> siteSet = new HashSet<>();
// last month
Date yesterday = monthDate(DateUtils.addDays(new Date(), -1)); // method that returns the date on the first of the month
Date startTime = startofYear(new Date(yesterday.getTime())); // method that returns the date on the first of the year
// list all the sites with a metric file
JavaPairRDD<String, String> allMetricFiles = jsc.wholeTextFiles("hdfs:///somePath/*/poc.json");
for ( Tuple2<String, String> each : allMetricFiles.toArray() ) {
logger.info("Reading from " + each._1);
DataFrame metric = sqlContext.read().format("json").load(each._1).cache();
metric.count();
boolean siteNameDisplayed = false;
boolean dateDisplayed = false;
do {
Date endTime = DateUtils.addMonths(startTime, 1);
HashSet<Row> totalUsersForThisMonth = new HashSet<>();
for (String dataPoint : Conf.DataPoints) { // This is a String[] with 4 elements for this specific case
try {
if (siteNameDisplayed == false) {
String siteName = parseSiteFromPath(each._1); // method returning a parsed String
logger.info("Data for site: " + siteName);
siteSet.add(siteName);
siteNameDisplayed = true;
}
if ( dateDisplayed == false ) {
logger.info("Month: " + formatDate(startTime)); // SimpleFormatDate("yyyy-MM-dd")
dateDisplayed = true;
}
DataFrame lastMonth = metric.filter("event.eventId=\"" + dataPoint + "\"").filter("creationDate >= " + startTime.getTime()).filter("creationDate < " + endTime.getTime()).select("event.data.UserId").distinct();
logger.info("Distinct for last month for " + dataPoint + ": " + lastMonth.count());
totalUsersForThisMonth.addAll(lastMonth.collectAsList());
} catch (Exception e) {
// data does not fit the expected model so there is nothing to print
}
}
logger.info("Total Unique for the month: " + totalStudentForThisMonth.size());
startTime = DateUtils.addMonths(startTime, 1);
dateDisplayed = false;
} while ( startTime.getTime() < commonTmsMetric.monthDate(yesterday).getTime());
// reset startTime for the next site
startTime = commonTmsMetric.StartofYear(new Date(yesterday.getTime()));
}
}
There are a few things that are not efficient in this code but when I look at the logs it only adds a few seconds to the whole processing.
I must be missing something big.
I have ran this with 2 executors and 1 executor and the difference is 20 seconds on 5 minutes.
This is running with Java 1.7 and Spark 1.4.1 on Hadoop 2.5.0.
Thank you!

So there a few things, but its hard to say without seeing the breakdown of the different tasks & their time. The short version is you are doing way to much work in the driver and not taking advantage of Spark's distributed capabilities.
For example, you are collecting all of the data back to the driver program (toArray() and your for loop). Instead you should just point Spark SQL at the files in needs to load.
For the operators, it seems like your doing many aggregations in the driver, instead you could use the driver to generate the aggregations and have Spark SQL execute them.
Another big difference between your in-house code and the DataFrame code is going to be Schema inference. Since you've already created classes to represent your data, it seems likely that you know the schema of your JSON data. You can likely speed up your code by adding the schema information at read time so Spark SQL can skip inference.
I'd suggest re-visiting this approach and trying to build something using Spark SQL's distributed operators.

Retrieve multiple messages from SQS

I have multiple messages in SQS. The following code always returns only one, even if there are dozens visible (not in flight). setMaxNumberOfMessages I thought would allow multiple to be consumed at once .. have i misunderstood this?
CreateQueueRequest createQueueRequest = new CreateQueueRequest().withQueueName(queueName);
String queueUrl = sqs.createQueue(createQueueRequest).getQueueUrl();
ReceiveMessageRequest receiveMessageRequest = new ReceiveMessageRequest(queueUrl);
receiveMessageRequest.setMaxNumberOfMessages(10);
List<Message> messages = sqs.receiveMessage(receiveMessageRequest).getMessages();
for (Message message : messages) {
// i'm a message from SQS
}
I've also tried using withMaxNumberOfMessages without any such luck:
receiveMessageRequest.withMaxNumberOfMessages(10);
How do I know there are messages in the queue? More than 1?
Set<String> attrs = new HashSet<String>();
attrs.add("ApproximateNumberOfMessages");
CreateQueueRequest createQueueRequest = new CreateQueueRequest().withQueueName(queueName);
GetQueueAttributesRequest a = new GetQueueAttributesRequest().withQueueUrl(sqs.createQueue(createQueueRequest).getQueueUrl()).withAttributeNames(attrs);
Map<String,String> result = sqs.getQueueAttributes(a).getAttributes();
int num = Integer.parseInt(result.get("ApproximateNumberOfMessages"));
The above always is run prior and gives me an int that is >1
Thanks for your input

AWS API Reference Guide: Query/QueryReceiveMessage
Due to the distributed nature of the queue, a weighted random set of machines is sampled on a ReceiveMessage call. That means only the messages on the sampled machines are returned. If the number of messages in the queue is small (less than 1000), it is likely you will get fewer messages than you requested per ReceiveMessage call. If the number of messages in the queue is extremely small, you might not receive any messages in a particular ReceiveMessage response; in which case you should repeat the request.
and
MaxNumberOfMessages: Maximum number of messages to return. SQS never returns more messages than this value but might return fewer.

There is a comprehensive explanation for this (arguably rather idiosyncratic) behaviour in the SQS reference documentation.
SQS stores copies of messages on multiple servers and receive message requests are made to these servers with one of two possible strategies,
Short Polling : The default behaviour, only a subset of the servers (based on a weighted random distribution) are queried.
Long Polling : Enabled by setting the WaitTimeSeconds attribute to a non-zero value, all of the servers are queried.
In practice, for my limited tests, I always seem to get one message with short polling just as you did.

I had the same problem. What is your Receive Message Wait Time for your queue set to? When mine was at 0, it only returned 1 message even if there were 8 in the queue. When I increased the Receive Message Wait Time, then I got all of them. Seems kind of buggy to me.

I was just trying the same and with the help of these two attributes setMaxNumberOfMessages and setWaitTimeSeconds i was able to get 10 messages.
ReceiveMessageRequest receiveMessageRequest = new ReceiveMessageRequest(myQueueUrl);
receiveMessageRequest.setMaxNumberOfMessages(10);
receiveMessageRequest.setWaitTimeSeconds(20);
Snapshot of o/p:
Receiving messages from TestQueue.
Number of messages:10
Message
MessageId: 31a7c669-1f0c-4bf1-b18b-c7fa31f4e82d
...

receiveMessageRequest.withMaxNumberOfMessages(10);
Just to be clear, the more practical use of this would be to add to your constructor like this:
ReceiveMessageRequest receiveMessageRequest = new ReceiveMessageRequest(queueUrl).withMaxNumberOfMessages(10);
Otherwise, you might as well just do:
receiveMessageRequest.setMaxNumberOfMessages(10);
That being said, changing this won't help the original problem.

Thanks Caoilte!
I faced this issue also. Finally solved by using long polling follow the configuration here:
https://docs.aws.amazon.com/AWSSimpleQueueService/latest/SQSDeveloperGuide/sqs-configure-long-polling-for-queue.html
Unfortunately, to use long polling, you must create your queue as FIFO one. I tried standard queue with no luck.
And when receiving, need also set MaxNumberOfMessages. So my code is like:
ReceiveMessageRequest receive_request = new ReceiveMessageRequest()
.withQueueUrl(QUEUE_URL)
.withWaitTimeSeconds(20)
.withMaxNumberOfMessages(10);
Although solved, still feel too wired. AWS should definitely provide a more neat API for this kind of basic receiving operation.
From my point, AWS has many many cool features but not good APIs. Like those guys are rushing out all the time.

For small task list I use FIFO queue like stackoverflow.com/a/55149351/13678017
for example modified AWS tutorial
// Create a queue.
System.out.println("Creating a new Amazon SQS FIFO queue called " + "MyFifoQueue.fifo.\n");
final Map<String, String> attributes = new HashMap<>();
// A FIFO queue must have the FifoQueue attribute set to true.
attributes.put("FifoQueue", "true");
/*
* If the user doesn't provide a MessageDeduplicationId, generate a
* MessageDeduplicationId based on the content.
*/
attributes.put("ContentBasedDeduplication", "true");
// The FIFO queue name must end with the .fifo suffix.
final CreateQueueRequest createQueueRequest = new CreateQueueRequest("MyFifoQueue4.fifo")
.withAttributes(attributes);
final String myQueueUrl = sqs.createQueue(createQueueRequest).getQueueUrl();
// List all queues.
System.out.println("Listing all queues in your account.\n");
for (final String queueUrl : sqs.listQueues().getQueueUrls()) {
System.out.println(" QueueUrl: " + queueUrl);
}
System.out.println();
// Send a message.
System.out.println("Sending a message to MyQueue.\n");
for (int i = 0; i < 4; i++) {
var request = new SendMessageRequest()
.withQueueUrl(myQueueUrl)
.withMessageBody("message " + i)
.withMessageGroupId("userId1");
;
sqs.sendMessage(request);
}
for (int i = 0; i < 6; i++) {
var request = new SendMessageRequest()
.withQueueUrl(myQueueUrl)
.withMessageBody("message " + i)
.withMessageGroupId("userId2");
;
sqs.sendMessage(request);
}
// Receive messages.
System.out.println("Receiving messages from MyQueue.\n");
var receiveMessageRequest = new ReceiveMessageRequest(myQueueUrl);
receiveMessageRequest.setMaxNumberOfMessages(10);
receiveMessageRequest.setWaitTimeSeconds(20);
// what receive?
receiveMessageRequest.withMessageAttributeNames("userId2");
final List<Message> messages = sqs.receiveMessage(receiveMessageRequest).getMessages();
for (final Message message : messages) {
System.out.println("Message");
System.out.println(" MessageId: "
+ message.getMessageId());
System.out.println(" ReceiptHandle: "
+ message.getReceiptHandle());
System.out.println(" MD5OfBody: "
+ message.getMD5OfBody());
System.out.println(" Body: "
+ message.getBody());
for (final Entry<String, String> entry : message.getAttributes()
.entrySet()) {
System.out.println("Attribute");
System.out.println(" Name: " + entry
.getKey());
System.out.println(" Value: " + entry
.getValue());
}
}

Here's a workaround, you can call receiveMessageFromSQS method asynchronously.
bulkReceiveFromSQS (queueUrl, totalMessages, asyncLimit, batchSize, visibilityTimeout, waitTime, callback) {
batchSize = Math.min(batchSize, 10);
let self = this,
noOfIterations = Math.ceil(totalMessages / batchSize);
async.timesLimit(noOfIterations, asyncLimit, function(n, next) {
self.receiveMessageFromSQS(queueUrl, batchSize, visibilityTimeout, waitTime,
function(err, result) {
if (err) {
return next(err);
}
return next(null, _.get(result, 'Messages'));
});
}, function (err, listOfMessages) {
if (err) {
return callback(err);
}
listOfMessages = _.flatten(listOfMessages).filter(Boolean);
return callback(null, listOfMessages);
});
}
It will return you an array with a given number of messages