My node application uses the kafka-node node module.
I have a kafka topic with three partitions as seen below:
Topic: NotifierTemporarye3df:/opPartitionCount: 3in$ kafReplicationFactor: 3ibe Configs: segment.bytes=1073741824 --topic NotifierTemporary
Topic: NotifierTemporary Partition: 0 Leader: 1001 Replicas: 1001,1003,1002 Isr: 1001,1003,1002
Topic: NotifierTemporary Partition: 1 Leader: 1002 Replicas: 1002,1001,1003 Isr: 1002,1001,1003
Topic: NotifierTemporary Partition: 2 Leader: 1003 Replicas: 1003,1002,1001 Isr: 1003,1002,1001
When I write a series of keyed messages to my topic, they all appear to be written to the same partition. I would expect some of my different keyed messages to be sent to partitions 1 and 2.
Here is my log output from the consumer onMessage event function for several messages:
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":66,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":345,"partition":0,"highWaterOffset":346,"key":"66","timestamp":"2020-03-19T00:16:57.783Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":222,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":346,"partition":0,"highWaterOffset":347,"key":"222","timestamp":"2020-03-19T00:16:57.786Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":13,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":347,"partition":0,"highWaterOffset":348,"key":"13","timestamp":"2020-03-19T00:16:57.791Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":316,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":348,"partition":0,"highWaterOffset":349,"key":"316","timestamp":"2020-03-19T00:16:57.798Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":446,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":349,"partition":0,"highWaterOffset":350,"key":"446","timestamp":"2020-03-19T00:16:57.806Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":66,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":350,"partition":0,"highWaterOffset":351,"key":"66","timestamp":"2020-03-19T00:17:27.918Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":222,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":351,"partition":0,"highWaterOffset":352,"key":"222","timestamp":"2020-03-19T00:17:27.920Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":13,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":352,"partition":0,"highWaterOffset":353,"key":"13","timestamp":"2020-03-19T00:17:27.929Z"}
the message is: {"topic":"NotifierTemporary","value":"{\"recipient\":316,\"subject\":\"download complete\",\"message\":\"s3/123.jpg\"}","offset":353,"partition":0,"highWaterOffset":354,"key":"316","timestamp":"2020-03-19T00:17:27.936Z"}
Here is the kafka-node producer code to send a message:
* #description Adds a notification message to the Kafka topic that is not saved in a database.
* #param {Int} recipientId - accountId of recipient of notification message
* #param {Object} message - message payload to send to recipient
*/
async sendTemporaryNotification(recipientId, subject, message) {
const notificationMessage = {
recipient: recipientId,
subject,
message,
};
// we need to validate this message schema - this will throw if invalid
Joi.assert(notificationMessage, NotificationMessage);
// partition based on the recipient
const payloads = [
{ topic: KAFKA_TOPIC_TEMPORARY, messages: JSON.stringify(notificationMessage), key: notificationMessage.recipient },
];
if (this.isReady) {
await this.producer.sendAsync(payloads);
}
else {
throw new ProducerNotReadyError('Notifier Producer not ready');
}
}
}
As you can see, none of them are ever from partitions 1 & 2. This is true even after constantly sending messages with random integer keys for several minutes. What could I be doing wrong?
The correct partitionerType needs to be configured when you create the producer:
// Partitioner type (default = 0, random = 1, cyclic = 2, keyed = 3, custom = 4), default 0
new Producer(client, {paritionerType: 3});
See the docs: https://www.npmjs.com/package/kafka-node#producerkafkaclient-options-custompartitioner
Scarysize was correct about me not specifying the partitioner type. For anyone wondering what a complete partitioned producer looks like, you can reference this code. I've verified that this distributes messages based off of the provided keys. I used a HighLevelProducer here because one of the main contributors of the kafka-node library suggested that others use it in order to solve partitioning issues. I have not verified that this solution would work with a regular Producer rather than the HighLevelProducer.
In this example, I'm sending notification messages to users based off of their userId. That is the key which messages are being partitioned on.
const { KafkaClient, HighLevelProducer, KeyedMessage } = require('kafka-node');
const Promise = require('bluebird');
const NotificationMessage = require(__dirname + '/../models/notificationMessage.js');
const ProducerNotReadyError = require(__dirname + '/../errors/producerNotReadyError.js');
const Joi = require('#hapi/joi');
const KAFKA_TOPIC_TEMPORARY = 'NotifierTemporary';
/**
* #classdesc Producer that sends notification messages to Kafka.
* #class
*/
class NotifierProducer {
/**
* Create NotifierProducer.
* #constructor
* #param {String} kafkaHost - address of kafka server
*/
constructor(kafkaHost) {
const client = Promise.promisifyAll(new KafkaClient({kafkaHost}));
const producerOptions = {
partitionerType: HighLevelProducer.PARTITIONER_TYPES.keyed, // this is a keyed partitioner
};
this.producer = Promise.promisifyAll(new HighLevelProducer(client, producerOptions));
this.isReady = false;
this.producer.on('ready', async () => {
await client.refreshMetadataAsync([KAFKA_TOPIC_TEMPORARY]);
console.log('Notifier Producer is operational');
this.isReady = true;
});
this.producer.on('error', err => {
console.error('Notifier Producer error: ', err);
this.isReady = false;
});
}
/**
* #description Adds a notification message to the Kafka topic that is not saved in a database.
* #param {Int} recipientId - accountId of recipient of notification message
* #param {String} subject - subject header of the message
* #param {Object} message - message payload to send to recipient
*/
async sendTemporaryNotification(recipientId, subject, message) {
const notificationMessage = {
recipient: recipientId,
subject,
message,
};
// we need to validate this message schema - this will throw if invalid
Joi.assert(notificationMessage, NotificationMessage);
// partition based on the recipient
const messageKM = new KeyedMessage(notificationMessage.recipient, JSON.stringify(notificationMessage));
const payloads = [
{ topic: KAFKA_TOPIC_TEMPORARY, messages: messageKM, key: notificationMessage.recipient },
];
if (this.isReady) {
await this.producer.sendAsync(payloads);
}
else {
throw new ProducerNotReadyError('Notifier Producer not ready');
}
}
}
/**
* Kafka topic that the producer and corresponding consumer will use to send temporary messages.
* #type {string}
*/
NotifierProducer.KAFKA_TOPIC_TEMPORARY = KAFKA_TOPIC_TEMPORARY;
module.exports = NotifierProducer;
Related
I am trying to connect to kafka server. Authentication is based on GSSAPI.
/opt/app-root/src/server/node_modules/node-rdkafka/lib/error.js:411
return new LibrdKafkaError(e);
^
Error: broker transport failure
at Function.createLibrdkafkaError (/opt/app-root/src/server/node_modules/node-rdkafka/lib/error.js:411:10)
at /opt/app-root/src/server/node_modules/node-rdkafka/lib/client.js:350:28
This my test_kafka.js:
const Kafka = require('node-rdkafka');
const kafkaConf = {
'group.id': 'espdev2',
'enable.auto.commit': true,
'metadata.broker.list': 'br01',
'security.protocol': 'SASL_SSL',
'sasl.kerberos.service.name': 'kafka',
'sasl.kerberos.keytab': 'svc_esp_kafka_nonprod.keytab',
'sasl.kerberos.principal': 'svc_esp_kafka_nonprod#INT.LOCAL',
'debug': 'all',
'enable.ssl.certificate.verification': true,
//'ssl.certificate.location': 'some-root-ca.cer',
'ssl.ca.location': 'some-root-ca.cer',
//'ssl.key.location': 'svc_esp_kafka_nonprod.keytab',
};
const topics = 'hello1';
console.log(Kafka.features);
let readStream = new Kafka.KafkaConsumer.createReadStream(kafkaConf, { "auto.offset.reset": "earliest" }, { topics })
readStream.on('data', function (message) {
const messageString = message.value.toString();
console.log(`Consumed message on Stream: ${messageString}`);
});
You can look at this issue for the explanation of this error:
https://github.com/edenhill/librdkafka/issues/1987
Taken from #edenhill:
As a general rule for librdkafka-based clients: given that the cluster and client are correctly configured, all errors can be ignored as they are most likely temporary and librdkafka will attempt to recover automatically. In this specific case; if a group coordinator request fails it will be retried (using any broker in state Up) within 500ms. The current assignment and group membership will not be affected, if a new coordinator is found before the missing heartbeats times out the membership (session.timeout.ms).
Auto offset commits will be stalled until a new coordinator is found. In a future version we'll extend the error type to include a severity, allowing applications to happily ignore non-terminal errors. At this time an application should consider all errors informational, and not terminal.
My subscriber looks like this:
from google.cloud import pubsub_v1
from google.cloud.pubsub_v1.types import DeadLetterPolicy
dead_letter_policy = DeadLetterPolicy(
dead_letter_topic='dead_letter_topic',
max_delivery_attempts=5,
)
topic_path = subscriber.topic_path(PROJECT, TOPIC)
subscriber.create_subscription(sub_path, topic_path, dead_letter_policy=dead_letter_policy)
subscriber = pubsub_v1.SubscriberClient()
subscription_path = subscriber.subscription_path(PROJECT, SUBSCRIPTION)
def callback(message):
print("Received message: {}".format(message))
print('Attempted:', message.delivery_attempt, 'times')
data = message.data.decode('utf-8')
data_d = json.loads(data)
if data_d["name"] == "some_file.json":
message.nack()
else:
message.ack()
The received message looks like this:
Received message: Message {
data: b'{\n "kind": "storage#object",\n "id": "...'
attributes: {
"bucketId": "sample_bucket",
...
}
}
Attempted: 12 times
Clearly it attempted more than 5 times, why I can still pull this message from PubSub topic?
Here is the subscription info:
ackDeadlineSeconds: 10
deadLetterPolicy:
deadLetterTopic: projects/someproject/topics/dead_letter
maxDeliveryAttempts: 5
expirationPolicy:
ttl: 2678400s
messageRetentionDuration: 604800s
name: projects/someproject/subscriptions/new_sub
pushConfig: {}
topic: projects/someproject/topics/pubsub_sample
Typically, this happens if you haven't given Pub/Sub permission to publish to your dead letter topic or subscribe to your subscription. You need to ensure you have run the following:
PUBSUB_SERVICE_ACCOUNT="service-${PROJECT_NUMBER}#gcp-sa-pubsub.iam.gserviceaccount.com"
gcloud pubsub topics add-iam-policy-binding <dead letter topic> \
--member="serviceAccount:${PUBSUB_SERVICE_ACCOUNT}"\
--role='roles/pubsub.publisher'
gcloud pubsub subscriptions add-iam-policy-binding <subscription with dead letter queue> \
--member="serviceAccount:${PUBSUB_SERVICE_ACCOUNT}"\
--role='roles/pubsub.subscriber'
If writing to the dead letter queue topic fails, then Cloud Pub/Sub will continue to deliver the message to your subscriber.
I have a created a HighLevelProducer to publish messages to a topic stream that will be consumed by a ConsumerGroupStream using kafka-node. When I create multiple consumers from the same ConsumerGroup to consume from that same topic only one partition is created and only one consumer is consuming. I have also tried to define the number of partitions for that topic although I'm not sure if is required to define it upon creating the topic and if so how many partitions will I need in advance. In addition, is it possible to push an object to the Transform stream and not a string (I currently used JSON.stringify because otherwise I got [Object object] in the consumer.
const myProducerStream = ({ kafkaHost, highWaterMark, topic }) => {
const kafkaClient = new KafkaClient({ kafkaHost });
const producer = new HighLevelProducer(kafkaClient);
const options = {
highWaterMark,
kafkaClient,
producer
};
kafkaClient.refreshMetadata([topic], err => {
if (err) throw err;
});
return new ProducerStream(options);
};
const transfrom = topic => new Transform({
objectMode: true,
decodeStrings: true,
transform(obj, encoding, cb) {
console.log(`pushing message ${JSON.stringify(obj)} to topic "${topic}"`);
cb(null, {
topic,
messages: JSON.stringify(obj)
});
}
});
const publisher = (topic, kafkaHost, highWaterMark) => {
const myTransfrom = transfrom(topic);
const producer = myProducerStream({ kafkaHost, highWaterMark, topic });
myTransfrom.pipe(producer);
return myTransform;
};
The consumer:
const createConsumerStream = (sourceTopic, kafkaHost, groupId) => {
const consumerOptions = {
kafkaHost,
groupId,
protocol: ['roundrobin'],
encoding: 'utf8',
id: uuidv4(),
fromOffset: 'latest',
outOfRangeOffset: 'earliest',
};
const consumerGroupStream = new ConsumerGroupStream(consumerOptions, sourceTopic);
consumerGroupStream.on('connect', () => {
console.log(`Consumer id: "${consumerOptions.id}" is connected!`);
});
consumerGroupStream.on('error', (err) => {
console.error(`Consumer id: "${consumerOptions.id}" encountered an error: ${err}`);
});
return consumerGroupStream;
};
const publisher = (func, destTopic, consumerGroupStream, kafkaHost, highWaterMark) => {
const messageTransform = new AsyncMessageTransform(func, destTopic);
const resultProducerStream = myProducerStream({ kafkaHost, highWaterMark, topic: destTopic })
consumerGroupStream.pipe(messageTransform).pipe(resultProducerStream);
};
For the first question:
The maximum working consumers in a group are equal to the number of partitions.
So if you have TopicA with 1 partition and you have 5 consumers in your consumer group, 4 of them will be idle.
If you have TopicA with 5 partitions and you have 5 consumers in your consumer group, all of them will be active and consuming messages from your topic.
To specify the number of partitions, you should create the topic from CLI instead of expecting Kafka to create it when you first publish messages.
To create a topic with a specific number of partitions:
bin/kafka-topics.sh --create --bootstrap-server localhost:9092 --replication-factor 3 --partitions 3 --topic test
To alter the number of partitions in an already existed topic:
bin/kafka-topics.sh --zookeeper zk_host:port/chroot --alter --topic test
--partitions 40
Please note that you can only increase the number of partitions, you can not decrease them.
Please refer to Kafka Docs https://kafka.apache.org/documentation.html
Also if you'd like to understand more about Kafka please check the free book https://www.confluent.io/resources/kafka-the-definitive-guide/
I have been seeing so many kafka consumer rebalances even if the thread is consuming nothing.I would expect the consumer not to rebalance in this scenario.
Here is the sample code.
import argparse
from kafka.coordinator.assignors.range import RangePartitionAssignor
from kafka import KafkaConsumer
import time
import sys
import logging
from kafka.consumer.subscription_state import ConsumerRebalanceListener
logger = logging.getLogger()
logger.setLevel(logging.DEBUG)
ch = logging.StreamHandler(sys.stdout)
ch.setLevel(logging.DEBUG)
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
ch.setFormatter(formatter)
logger.addHandler(ch)
def get_config(args):
config = {
'bootstrap_servers': args.host,
'group_id': args.group,
'key_deserializer': lambda msg: msg,
'value_deserializer': lambda msg: msg,
'partition_assignment_strategy': [RangePartitionAssignor],
'max_poll_records': args.records,
'auto_offset_reset': args.offset,
# 'max_poll_interval_ms': 300000,
# 'connections_max_idle_ms': 8 * 60 * 1000
}
return config
def start_consumer(args):
config = get_config(args)
consumer = KafkaConsumer(**config)
consumer.subscribe([args.topic],
listener=RepartitionListener(None))
for record in consumer:
print record.offset, record.partition
time.sleep(int(args.delay) / 1000.0)
class RepartitionListener(ConsumerRebalanceListener):
def __init__(self):
pass
def on_partitions_revoked(self, revoked):
print("partition revoked ")
for tp in revoked:
try:
print("[{}] revoked topic = {} partition = {}".format(time.strftime("%c"),
tp.topic, tp.partition))
partition_key = "{}_{}".format(tp.topic, str(tp.partition))
except Exception as e:
print("Got exception partition_key = {} {}".
format(tp, e.message))
def on_partitions_assigned(self, assigned):
pass
def main():
parser = argparse.ArgumentParser(
description='Tool to test consumer group with delay')
named_args = parser.add_argument_group('snamed arguments')
named_args.add_argument('-g', '--group', help='group id for the consumer',
required=True)
named_args.add_argument('-r', '--records', help='num records to consume',
required=True)
named_args.add_argument('-k', '--topic', help='kafka topic', required=True)
named_args.add_argument('-d', '--delay', help='add process delay in ms', required=True)
named_args.add_argument('-s', '--host', help='Kafka host format host:port', required=False)
parser.add_argument('-o', '--offset',
default='latest',
help='offset to read from earliest/latest')
args = parser.parse_args()
print args
start_consumer(args)
if __name__ == "__main__":
main()
How to avoid the rebalance trigger? From the logs I am seeing heartbeat is failing, But I am expecting the heartbeat to be continued even if there is no messages over a period of time greater than session.time.out.ms.
2019-02-27 20:39:43,281 - kafka.coordinator - WARNING - Heartbeat session expired, marking coordinator dead
I have two channels in my application which hare bound to two Kafka topics:
input
error.input.my-group
Input is configured in order to send message to dlq (error.input.my-group) in case of error.
I have a StreamListener on "error.input.my-group" which is configured in order to send the message back to original channel.
#StreamListener(Channels.DLQ)
#SendTo(Channels.INPUT)
public Message<?> reRoute(Message<?> failed){
messageDeliveryService.waitUntilCanBeDelivered(failed);
processed.incrementAndGet();
Integer retries = failed.getHeaders().get(X_RETRIES_HEADER, Integer.class);
retries = retries == null ? 1 : retries+1;
if (retries < MAX_RETRIES) {
logger.info("Retry (count={}) for {}", retries, failed);
return buildRetryMessage(failed, retries);
}
else {
logger.error("Retries exhausted (-> sent to parking lot) for {}", failed);
Channels.parkingLot().send(MessageBuilder.fromMessage(failed)
.setHeader(BinderHeaders.PARTITION_OVERRIDE,
failed.getHeaders().get(KafkaHeaders.RECEIVED_PARTITION_ID))
.build());
}
return null;
}
private Message<?> buildRetryMessage(Message<?> failed, int retries) {
return MessageBuilder.fromMessage(failed)
.setHeader(X_RETRIES_HEADER, retries)
.setHeader(BinderHeaders.PARTITION_OVERRIDE,
failed.getHeaders().get(KafkaHeaders.RECEIVED_PARTITION_ID))
.build();
}
Here is my Channels class
#Component
public interface Channels {
String INPUT = "INPUT";
//Default name use by SCS (error.<input-topic-name>.<group-name>)
String DLQ = "error.input.my-group";
String PARKING_LOT = "parkingLot.input.my-group";
#Input(INPUT)
SubscribableChannel input();
#Input(DLQ)
SubscribableChannel dlq();
#Output(PARKING_LOT)
MessageChannel parkingLot();
}
Here is my configuration
spring:
cloud:
stream:
default:
group: my-group
binder:
headerMode: headers kafka:
binder:
# Necessary in order to commit the message to all the Kafka brokers handling the partition -> maximum durability
# -1 = all
requiredAcks: -1
brokers: bootstrap.kafka.svc.cluster.local:9092,bootstrap.kafka.svc.cluster.local:9093,bootstrap.kafka.svc.cluster.local:9094,bootstrap.kafka.svc.cluster.local:9095,bootstrap.kafka.svc.cluster.local:9096,bootstrap.kafka.svc.cluster.local:9097
bindings:
input:
consumer:
partitioned: true
enableDlq: true
dlqProducerProperties:
configuration:
key.serializer: "org.apache.kafka.common.serialization.ByteArraySerializer"
"[error.input.my-group]":
consumer:
# We cannot loose any message and we don't have any DLQ for the DLQ, therefore we only commit in case of success
autoCommitOnError: false
ackEachRecord: true
partitioned: true
enableDlq: false
bindings:
input:
contentType: application/xml
destination: input
"[error.input.my-group]":
contentType: application/xml
destination: error.input.my-group
"[parkingLot.input.my-group]":
contentType: application/xml
destination: parkingLot.input.my-group
Problem is my messages are never pushed again to Kafka but directly delivered to my input channel. Is there something I misunderstood?
In order to #SendTo the kafka destination instead of directly, you need an output binding.