How can I send large messages with Kafka (over 15MB)? - java

I send String-messages to Kafka V. 0.8 with the Java Producer API.
If the message size is about 15 MB I get a MessageSizeTooLargeException.
I have tried to set message.max.bytesto 40 MB, but I still get the exception. Small messages worked without problems.
(The exception appear in the producer, I don't have a consumer in this application.)
What can I do to get rid of this exception?
My example producer config
private ProducerConfig kafkaConfig() {
Properties props = new Properties();
props.put("metadata.broker.list", BROKERS);
props.put("serializer.class", "kafka.serializer.StringEncoder");
props.put("request.required.acks", "1");
props.put("message.max.bytes", "" + 1024 * 1024 * 40);
return new ProducerConfig(props);
}
Error-Log:
4709 [main] WARN kafka.producer.async.DefaultEventHandler - Produce request with correlation id 214 failed due to [datasift,0]: kafka.common.MessageSizeTooLargeException
4869 [main] WARN kafka.producer.async.DefaultEventHandler - Produce request with correlation id 217 failed due to [datasift,0]: kafka.common.MessageSizeTooLargeException
5035 [main] WARN kafka.producer.async.DefaultEventHandler - Produce request with correlation id 220 failed due to [datasift,0]: kafka.common.MessageSizeTooLargeException
5198 [main] WARN kafka.producer.async.DefaultEventHandler - Produce request with correlation id 223 failed due to [datasift,0]: kafka.common.MessageSizeTooLargeException
5305 [main] ERROR kafka.producer.async.DefaultEventHandler - Failed to send requests for topics datasift with correlation ids in [213,224]
kafka.common.FailedToSendMessageException: Failed to send messages after 3 tries.
at kafka.producer.async.DefaultEventHandler.handle(Unknown Source)
at kafka.producer.Producer.send(Unknown Source)
at kafka.javaapi.producer.Producer.send(Unknown Source)

You need to adjust three (or four) properties:
Consumer side:fetch.message.max.bytes - this will determine the largest size of a message that can be fetched by the consumer.
Broker side: replica.fetch.max.bytes - this will allow for the replicas in the brokers to send messages within the cluster and make sure the messages are replicated correctly. If this is too small, then the message will never be replicated, and therefore, the consumer will never see the message because the message will never be committed (fully replicated).
Broker side: message.max.bytes - this is the largest size of the message that can be received by the broker from a producer.
Broker side (per topic): max.message.bytes - this is the largest size of the message the broker will allow to be appended to the topic. This size is validated pre-compression. (Defaults to broker's message.max.bytes.)
I found out the hard way about number 2 - you don't get ANY exceptions, messages, or warnings from Kafka, so be sure to consider this when you are sending large messages.

Minor changes required for Kafka 0.10 and the new consumer compared to laughing_man's answer:
Broker: No changes, you still need to increase properties message.max.bytes and replica.fetch.max.bytes. message.max.bytes has to be equal or smaller(*) than replica.fetch.max.bytes.
Producer: Increase max.request.size to send the larger message.
Consumer: Increase max.partition.fetch.bytes to receive larger messages.
(*) Read the comments to learn more about message.max.bytes<=replica.fetch.max.bytes

The answer from #laughing_man is quite accurate. But still, I wanted to give a recommendation which I learned from Kafka expert Stephane Maarek. We actively applied this solution in our live systems.
Kafka isn’t meant to handle large messages.
Your API should use cloud storage (for example, AWS S3) and simply push a reference to S3 to Kafka or any other message broker. You'll need to find a place to save your data, whether it can be a network drive or something else entirely, but it shouldn't be a message broker.
If you don't want to proceed with the recommended and reliable solution above,
The message max size is 1MB (the setting in your brokers is called message.max.bytes) Apache Kafka. If you really needed it badly, you could increase that size and make sure to increase the network buffers for your producers and consumers.
And if you really care about splitting your message, make sure each message split has the exact same key so that it gets pushed to the same partition, and your message content should report a “part id” so that your consumer can fully reconstruct the message.
If the message is text-based try to compress the data, which may reduce the data size, but not magically.
Again, you have to use an external system to store that data and just push an external reference to Kafka. That is a very common architecture and one you should go with and widely accepted.
Keep that in mind Kafka works best only if the messages are huge in amount but not in size.
Source: https://www.quora.com/How-do-I-send-Large-messages-80-MB-in-Kafka

The idea is to have equal size of message being sent from Kafka Producer to Kafka Broker and then received by Kafka Consumer i.e.
Kafka producer --> Kafka Broker --> Kafka Consumer
Suppose if the requirement is to send 15MB of message, then the Producer, the Broker and the Consumer, all three, needs to be in sync.
Kafka Producer sends 15 MB --> Kafka Broker Allows/Stores 15 MB --> Kafka Consumer receives 15 MB
The setting therefore should be:
a) on Broker:
message.max.bytes=15728640
replica.fetch.max.bytes=15728640
b) on Consumer:
fetch.message.max.bytes=15728640

You need to override the following properties:
Broker Configs($KAFKA_HOME/config/server.properties)
replica.fetch.max.bytes
message.max.bytes
Consumer Configs($KAFKA_HOME/config/consumer.properties)
This step didn't work for me. I add it to the consumer app and it was working fine
fetch.message.max.bytes
Restart the server.
look at this documentation for more info:
http://kafka.apache.org/08/configuration.html

I think, most of the answers here are kind of outdated or not entirely complete.
To refer on the answer of Sacha Vetter (with the update for Kafka 0.10), I'd like to provide some additional Information and links to the official documentation.
Producer Configuration:
max.request.size (Link) has to be increased for files bigger than 1 MB, otherwise they are rejected
Broker/Topic configuration:
message.max.bytes (Link) may be set, if one like to increase the message size on broker level. But, from the documentation: "This can be set per topic with the topic level max.message.bytes config."
max.message.bytes (Link) may be increased, if only one topic should be able to accept lager files. The broker configuration must not be changed.
I'd always prefer a topic-restricted configuration, due to the fact, that I can configure the topic by myself as a client for the Kafka cluster (e.g. with the admin client). I may not have any influence on the broker configuration itself.
In the answers from above, some more configurations are mentioned as necessary:
replica.fetch.max.bytes (Link) (Broker config)
From the documentation: "This is not an absolute maximum, if the first record batch in the first non-empty partition of the fetch is larger than this value, the record batch will still be returned to ensure that progress can be made."
max.partition.fetch.bytes (Link) (Consumer config)
From the documentation: "Records are fetched in batches by the consumer. If the first record batch in the first non-empty partition of the fetch is larger than this limit, the batch will still be returned to ensure that the consumer can make progress."
fetch.max.bytes (Link) (Consumer config; not mentioned above, but same category)
From the documentation: "Records are fetched in batches by the consumer, and if the first record batch in the first non-empty partition of the fetch is larger than this value, the record batch will still be returned to ensure that the consumer can make progress."
Conclusion: The configurations regarding fetching messages are not necessary to change for processing messages, lager than the default values of these configuration (had this tested in a small setup). Probably, the consumer may always get batches of size 1. However, two of the configurations from the first block has to be set, as mentioned in the answers before.
This clarification should not tell anything about performance and should not be a recommendation to set or not to set these configuration. The best values has to be evaluated individually depending on the concrete planned throughput and data structure.

One key thing to remember that message.max.bytes attribute must be in sync with the consumer's fetch.message.max.bytes property. the fetch size must be at least as large as the maximum message size otherwise there could be situation where producers can send messages larger than the consumer can consume/fetch. It might worth taking a look at it.
Which version of Kafka you are using? Also provide some more details trace that you are getting. is there some thing like ... payload size of xxxx larger
than 1000000 coming up in the log?

For people using landoop kafka:
You can pass the config values in the environment variables like:
docker run -d --rm -p 2181:2181 -p 3030:3030 -p 8081-8083:8081-8083 -p 9581-9585:9581-9585 -p 9092:9092
-e KAFKA_TOPIC_MAX_MESSAGE_BYTES=15728640 -e KAFKA_REPLICA_FETCH_MAX_BYTES=15728640 landoop/fast-data-dev:latest `
This sets topic.max.message.bytes and replica.fetch.max.bytes on the broker.
And if you're using rdkafka then pass the message.max.bytes in the producer config like:
const producer = new Kafka.Producer({
'metadata.broker.list': 'localhost:9092',
'message.max.bytes': '15728640',
'dr_cb': true
});
Similarly, for the consumer,
const kafkaConf = {
"group.id": "librd-test",
"fetch.message.max.bytes":"15728640",
... .. }

Here is how I achieved successfully sending data up to 100mb using kafka-python==2.0.2:
Broker:
consumer = KafkaConsumer(
...
max_partition_fetch_bytes=max_bytes,
fetch_max_bytes=max_bytes,
)
Producer (See final solution at the end):
producer = KafkaProducer(
...
max_request_size=KafkaSettings.MAX_BYTES,
)
Then:
producer.send(topic, value=data).get()
After sending data like this, the following exception appeared:
MessageSizeTooLargeError: The message is n bytes when serialized which is larger than the total memory buffer you have configured with the buffer_memory configuration.
Finally I increased buffer_memory (default 32mb) to receive the message on the other end.
producer = KafkaProducer(
...
max_request_size=KafkaSettings.MAX_BYTES,
buffer_memory=KafkaSettings.MAX_BYTES * 3,
)

Related

How to avoid losing messages with Kafka streams

We have a streams application that consumes messages from a source topic, does some processing and forward the results to a destination topic.
The structure of the messages are controlled by some avro schemas.
When starting consuming messages if the schema is not cached yet the application will try to retrieve it from schema registry. If for whichever reason the schema registry is not available (say a network glitch) then the currently being processed message is lost because the default handler is something called LogAndContinueExceptionHandler.
o.a.k.s.e.LogAndContinueExceptionHandler : Exception caught during Deserialization, taskId: 1_5, topic: my.topic.v1, partition: 5, offset: 142768
org.apache.kafka.common.errors.SerializationException: Error retrieving Avro schema for id 62
Caused by: java.net.SocketTimeoutException: connect timed out
at java.base/java.net.PlainSocketImpl.socketConnect(Native Method) ~[na:na]
...
o.a.k.s.p.internals.RecordDeserializer : stream-thread [my-app-StreamThread-3] task [1_5] Skipping record due to deserialization error. topic=[my.topic.v1] partition=[5] offset=[142768]
...
So my question is what would be the proper way of dealing with situations like described above and make sure you don't lose messages no matter what. Is there an out of the box LogAndRollbackExceptionHandler error handler or a way of implementing your own?
Thank you in advance for your inputs.
I've not worked a lot on Kafka, but when i did, i remember having issues such as the one you are describing in our system.
Let me tell you how we took care of our scenarios, maybe it would help you out too:
Scenario 1: If your messages are being lost at the publishing side (publisher --> kafka), you can configure Kafka acknowledgement setting according to your need, if you use spring cloud stream with kafka, the property is spring.cloud.stream.kafka.binder.required-acks
Possible values:
At most once (Ack=0)
Publisher does not care if Kafka acknowledges or not.
Send and forget
Data loss is possible
At least once (Ack=1)
If Kafka does not acknowledge, publisher resends message.
Possible duplication.
Acknowledgment is sent before message is copied to replicas.
Exactly once (Ack=all)
If Kafka does not acknowledge, publisher resends message.
However, if a message gets sent more than once to Kafka, there is no duplication.
Internal sequence number, used to decide if message has already been written on topic or not.
Min.insync.replicas property needs to be set to ensure what is the minimum number of replices that need to be synced before kafka acknowledges to the producer.
Scenario 2: If your data is being lost at the consumer side (kafka --> consumer), you can change the Auto Commit feature of Kafka according to your usage. This is the property if you are using Spring cloud stream spring.cloud.stream.kafka.bindings.input.consumer.AutoCommitOffset.
By default, AutoCommitOffset is true in kafka, and every message that is sent to the consumer is "committed" at Kafka's end, meaning it wont be sent again. However if you change AutoCommitOffset to false, you will have the power to poll the message from kafka in your code, and once you are done with your work, explicitly set commit to true to let kafka know that now you are done with the message.
If a message is not committed, kafka will keep resending it until it is.
Hope this helps you out, or atleast points you in the right direction.

Kafka producer fails to send messages with NOT_LEADER_FOR_PARTITION exception

We're using spring-cloud-stream-binder-kafka (3.0.3.RELEASE) to send messages to our Kafka cluster (2.4.1). Every now and then one of the producer threads receives NOT_LEADER_FOR_PARTITION exceptions, and even exceeds the retries (currently set at 12, activated by dependency spring-retry). We've restricted the retries because we're sending about 1k msg/s (per producer instance) and were worried about the size of the buffer.
This way we're regularly loosing messages, which is bad for downstream consumers, because we can't simply reproduce the incoming traffic.
The error message is
[Producer clientId=producer-5] Received invalid metadata error in produce request on partition topic-21 due to org.apache.kafka.common.errors.NotLeaderForPartitionException: This server is not the leader for that topic-partition.. Going to request metadata update now
[Producer clientId=producer-5] Got error produce response with correlation id 974706 on topic-partition topic-21, retrying (8 attempts left). Error: NOT_LEADER_FOR_PARTITION
[Producer clientId=producer-5] Got error produce response with correlation id 974707 on topic-partition topic-21, retrying (1 attempts left). Error: NOT_LEADER_FOR_PARTITION
Any known way to avoid this?
Should we go back to the default of MAX_INT retries?
Why does it keep sending to the same broker, even though it responded with NOT_LEADER_FOR_PARTITION?
Any hints are welcome.
EDIT: We just noticed that the broker metric kafka_network_requestmetrics_responsequeuetimems goes up around that time, but the max we've seen is around 2.5s
Both Produce and Fetch requests are send to the leader replica of the partition. NotLeaderForPartitionException the exception is thrown when the request is sent to the partition which not the leader replica of the partition now.
The client maintains the information regarding the leader of each partition as a cache. The complete process of cache management is shown below.
The client needs to refresh this information by setting the metadata.max.age.ms in producer configuration. The default value of this tag is 300000 ms
You can go through the following Apache Kafka documentation.
https://kafka.apache.org/documentation/
Please go through the Sender.java code.
https://github.com/a0x8o/kafka/blob/master/clients/src/main/java/org/apache/kafka/clients/producer/internals/Sender.java
You will find both the error messages in the sender code. The default value of metadata.max.age.ms is 3 seconds. I think you should reduce this value and then observe the behavior.
you need config listeners properly
I'm using docker-compose like
services:
zookeeper:
container_name: zookeeper
ports:
- "2181:2181"
...
broker-1:
hostname: "broker-1.mydomain.com"
ports:
- "29091:29091"
...
broker-2:
hostname: "broker-2.mydomain.com"
container_name: broker-2
ports:
- "29092:29092"
edit server.properties for each broker
broker-1
listeners: PRIVATE_HOSTNAME://broker-1.mydomain.com:9092,PUBLIC_HOSTNAME://broker-1.mydomain.com:29091
advertised.listeners: PRIVATE_HOSTNAME://broker-1.mydomain.com:9092,PUBLIC_HOSTNAME://broker-1.mydomain.com:29091
listener.security.protocol.map: PUBLIC_HOSTNAME:PLAINTEXT,PRIVATE_HOSTNAME:PLAINTEXT
inter.broker.listener.name: PRIVATE_HOSTNAME
broker-2
listeners: PRIVATE_HOSTNAME://broker-2.mydomain.com:9092, PUBLIC_HOSTNAME://broker-2.mydomain.com:29092
advertised.listeners: PRIVATE_HOSTNAME://broker-2.mydomain.com:9092, PUBLIC_HOSTNAME://broker-2.mydomain.com:29092
listener.security.protocol.map: PUBLIC_HOSTNAME:PLAINTEXT, PRIVATE_HOSTNAME:PLAINTEXT
inter.broker.listener.name: PRIVATE_HOSTNAME
IMPORTANT: note that I'm using the same hostname for private and public net, because the consumer/producer can only access to kafka by register name like this:
[BrokerToControllerChannelManager broker=1 name=forwarding]: Recorded new controller, from now on will use broker broker-1.mydomain.com:9092
...
[BrokerToControllerChannelManager broker=2 name=forwarding]: Recorded new controller, from now on will use broker broker-2.mydomain.com:9092
edit your host /etc/hosts
127.0.0.1 broker-1.mydomain.com
127.0.0.1 broker-2.mydomain.com

Kafka only subscribe to latest message?

Sometimes(seems very random) Kafka sends old messages. I only want the latest messages so it overwrite messages with the same key. Currently it looks like I have multiple messages with the same key it doesn't get compacted.
I use this setting in the topic:
cleanup.policy=compact
I'm using Java/Kotlin and Apache Kafka 1.1.1 client.
Properties(8).apply {
val jaasTemplate = "org.apache.kafka.common.security.scram.ScramLoginModule required username=\"%s\" password=\"%s\";"
val jaasCfg = String.format(jaasTemplate, Configuration.kafkaUsername, Configuration.kafkaPassword)
put(ConsumerConfig.BOOTSTRAP_SERVERS_CONFIG,
BOOTSTRAP_SERVERS)
put(ConsumerConfig.GROUP_ID_CONFIG,
"ApiKafkaKotlinConsumer${Configuration.kafkaGroupId}")
put(ConsumerConfig.KEY_DESERIALIZER_CLASS_CONFIG,
StringDeserializer::class.java.name)
put(ConsumerConfig.VALUE_DESERIALIZER_CLASS_CONFIG,
StringDeserializer::class.java.name)
put("security.protocol", "SASL_SSL")
put("sasl.mechanism", "SCRAM-SHA-256")
put("sasl.jaas.config", jaasCfg)
put("max.poll.records", 100)
put("receive.buffer.bytes", 1000000)
}
Have I missed some settings?
If you want have only one value for each key, you have to use KTable<K,V> abstraction: StreamsBuilder::table(final String topic) from Kafka Streams. Topic used here should have cleanup policy set to compact.
If you use KafkaConsumer you just pull data from brokers. It doesn't give you any mechanism that perform some kind of deduplication. Depending on if compaction was performed or not, you can get one to n messages for same key.
Regarding compaction
Compaction doesn't mean, that all old value for same key are removed immediately. When old message for same key will be removed, depends on several properties. The most important are:
log.cleaner.min.cleanable.ratio
The minimum ratio of dirty log to total log for a log to eligible for cleaning
log.cleaner.min.compaction.lag.ms
The minimum time a message will remain uncompacted in the log. Only applicable for logs that are being compacted.
log.cleaner.enable
Enable the log cleaner process to run on the server. Should be enabled if using any topics with a cleanup.policy=compact including the internal offsets topic. If disabled those topics will not be compacted and continually grow in size.
More detail about compaction you can find https://kafka.apache.org/documentation/#compaction

Spring Kafka producers throwing TimeoutExceptions

Problem
I have a Kafka setup with three brokers in Kubernetes, set up according to the guide at https://github.com/Yolean/kubernetes-kafka. The following error message appears when producing messages from a Java client.
2018-06-06 11:15:44.103 ERROR 1 --- [ad | producer-1] o.s.k.support.LoggingProducerListener : Exception thrown when sending a message with key='null' and payload='[...redacted...]':
org.apache.kafka.common.errors.TimeoutException: Expiring 1 record(s) for topicname-0: 30001 ms has passed since last append
Detailed setup
The listeners are set up to allow SSL producers/consumers from the outside world:
advertised.host.name = null
advertised.listeners = OUTSIDE://kafka-0.mydomain.com:32400,PLAINTEXT://:9092
advertised.port = null
listener.security.protocol.map = PLAINTEXT:PLAINTEXT,SSL:SSL,SASL_PLAINTEXT:SASL_PLAINTEXT,SASL_SSL:SASL_SSL,OUTSIDE:SSL
listeners = OUTSIDE://:9094,PLAINTEXT://:9092
inter.broker.listener.name = PLAINTEXT
host.name =
port.name = 9092
The OUTSIDE listeners are listening on kafka-0.mydomain.com, kafka-1.mydomain.com, etc. The plaintext listeners are listening on any IP, since they are cluster-local to Kubernetes.
The producer settings:
kafka:
bootstrap-servers: kafka.mydomain.com:9092
properties:
security.protocol: SSL
producer:
batch-size: 16384
buffer-memory: 1048576 # 1MB
retries: 1
ssl:
key-password: redacted
keystore-location: file:/var/private/ssl/kafka.client.keystore.jks
keystore-password: redacted
truststore-location: file:/var/private/ssl/kafka.client.truststore.jks
truststore-password: redacted
In addition I set linger.ms to 100 in code, which forces messages to be transmitted within 100ms. Linger time is set intentionally low, because the use case requires minimal delays.
Analysis
The errors started appearing when the broker was moved moved to SSL.
On the server side everything is running as expected, there are no errors in the log and I can connect to the broker manually with a Kafka client tool.
The errors appear intermittently: sometimes it sends 30+ messages per second, sometimes it sends nothing at all. It may work like a charm for hours and then just spam timeouts for a little while.
Clocks for the client and server are in sync (UTC).
CPU is consistently around 20% for both the producing and server side.
What could it be?
This problem normally occurs when the producer is faster than the brokers, the reason why this happens with your setup seems to be that the SSL needs extra CPU and that may slow down the brokers. But anyway check the following:
Check if you are producing message at the same speed, according what you are saying seems that you are having spikes.
Another possibility is that other kafka clients in the cluster (producer or consumers), which not necessarily uses the same topic, makes this to happen because overloads the brokers (check brokers cpu/network).
To minimize whatever causes this retention you should increase the buffer-memory to more than 32MB, think that 32MB is the default and you are setting this lower. The lower you have, the easy is that the buffer gets full and if this happens it will block at most max.block.ms, and a request will timeout after request.timeout.ms.
Another parameter that you should increase is batch-size, this parameter is in bytes, not in number of messages. Also linger.ms should be increased, in case this producer messages are created in user request time, do not increase very much, a good choice could be 1-4 ms.
Messages will be send when the batch.size gets full or takes longer than linger.ms to have more data than the batch.size. Big batches increases the throughput in normal cases, but if the linger is too low it doesn't help, because you will send before you have enough data to get the batch.size.
Also recheck on producer logs that the properties are loaded correctly.

2 messages produced to same partition one by one - message 1 overridden by next as kafka producer (0.8.2.1) returns same offset

I have a unique problem which is happening like 50-100 times a day with message volume of ~2 millions per day on the topic.I am using Kafka producer API 0.8.2.1 and I have 12 brokers (v 0.8.2.2) running in prod with replication of 4. I have a topic with 60 partitions and I am calculating partition for all my messages and providing the value in the ProducerRecord itself. Now, the issue -
Application creates 'ProducerRecord' using -
new ProducerRecord<String, String>(topic, 30, null, message1);
providing topic, value message1 and partition 30. Then application call the send method and future is returned -
// null is for callback
Future<RecordMetadata> future = producer.send(producerRecord. null);
Now, app prints the offset and partition value by calling get on Future and then getting values from RecordMetadata - this is what i get -
Kafka Response : partition 30, offset 3416092
Now, the app produce the next message - message2 to same partition -
new ProducerRecord<String, String>(topic, 30, null, message2);
and kafka response -
Kafka Response : partition 30, offset 3416092
I receive the same offset again, and if I pull message from the offset of partition 30 using simple consumer, it ends up being the message2 which essentially mean i lost the message1.
Based on KafkaProducer documentation KafkaProducer, I am using a single producer instance (static instance shared) among 10 threads.
The producer is thread safe and should generally be shared among all threads for best performance.
I am using all default properties for producer (except max.request.size: 10000000), the message (String payload) size can be a few kbs to a 500 kbs. I am using ack value of 1.
What am i doing wrong here? Is there something I can look into or any producer property or server property I can tweak to make sure i don't lose any messages. I need some help here soon as I am losing some critical messages in production which is not good at all coz because of no exception its even hard to find out the message lost unless downstream process reports it.
EDIT:
The servers and client are now updated to kafka version 0.8.2.2. Also, the 10 app threads each use their own instance of kafka producer now. We are seeing better performance but still there is message loss.
Producer Properties:
value.serializer: org.apache.kafka.common.serialization.StringSerializer
key.serializer: org.apache.kafka.common.serialization.StringSerializer
bootstrap.servers: {SERVER VIP ENDPOINT}
acks: 1
batch.size: 204800
linger.ms: 10
send.buffer.bytes: 1048576
max.request.size: 10000000

Categories