we have setup a Kafka/Zookeeper Cluster consisting of 3 Brokers. We have one producer, sending messages to one specific Kafka topic and a few consumer groups reading from said topic. Those consumers perform a leader election via Zookeeper for themselves (independent from Kafka).
The versions used are:
Kafka: 0.9.0.1
Zookeeper: 3.4.6 (included in the Kafka-Package)
All processes are managed by Supervisor. So far, everything works just fine. What we tried now (for testing purposes) was to simply kill off all Zookeeper processes and see what happens.
As we expected, our consumer processes couldn't connect to Zookeeper anymore. But unexpectedly, the Kafka Brokers still worked. Our producer didn't complain at all and was still able to write into the topic. While I couldn't use kafka/bin/kafka-topics.sh or similar, since they all require a zookeeper-parameter, I could still see the actual size of the topic-log grow. After restarting the zookeeper processes, everything again worked just like before.
What we couldn't figure out is now... what actually happened there?
We thought, Kafka would require a working Zookeeper-Connection and we couldn't find any explanation for this behaviour online.
When you have one node of zookeeper, broker will not be able to contact zookeeper, after broker discovers zookeeper is not reachable, broker also will become unreachable. Hence the producer and consumer.
In case of producer it starts dropping(reject the record). In case of consumer it can happen that, the read record which is not ack'ed may end up processing again when broker is up and ready...
in case of 3node zk one node failure is acceptable as quorum is still satisfied... but cant afford the 2node failures which will lead to the above consequences...
Related
I'm testing Kafka on Linux, but I don't know what's wrong because the test results are different from what I understand.
Let me explain the setting.
Currently, three brokers were configured with kafka version 2.8.1 in centos7 using 9092, 9093, and 9094 ports, respectively.
In the case of producers, all three ports were connected to the bootstrap-server setting and then executed.
kafka-console-producer.bat --broker-list localhost:9092,localhost:9093,localhost:9094 --topic test
In the case of consumers, three were set up so that they could be attached to each of the three ports.
1. kafka-console-consumer.bat --bootstrap-serverlocalhost:9092 --topic test
2. kafka-console-consumer.bat --bootstrap-serverlocalhost:9093 --topic test
3. kafka-console-consumer.bat --bootstrap-serverlocalhost:9094 --topic test
If I were to explain what I understood here,
In the case of Kafka, the leader broker acts as a controller, and the follow brokers copy the leader broker's data.
If one of the follower brokers dies, a disconnection message simply appears on the consumer connected to the broker.Other brokers operate normally.
If a leader broker dies, one of the follow brokers will be changed to a leader broker, and the changed leader broker will act as a controller.
If I were to explain the problem,
If you kill a leader broker, check the describe option, and the other follow broker has changed to a leader, but both producers and consumers cannot find a new leader and fail.
Even if a broker running on 9092 ports kills the broker without being a leader, the producer and consumer will fail.
Question.
If the leader broker dies, should the producer and consumer also set up a new connection?
Am I misunderstanding the producer and consumer?
Is there anything wrong with the setting?
I'm testing Kafka on Linux
But you're using Batch files, and connecting to localhost, which are for Windows...
so that they could be attached to each of the three ports.
This isn't how Kafka distributes load. You can only have one consumer thread active per topic partition. Unclear how many partitions your topic has, but if you have only one and that specific broker died (it is the only replica and leader), this explains why your clients would stop working.
Besides that, Kafka is generally on the same port, on mulitple hosts. Using one host is not truly fault-tolerant, and is a waste of resources (CPU, RAM, and disk).
Regarding producers, there is a property for retries that can be configured; I'm not sure if the console producer overrides the default or not, but it should connect to the next available broker upon a new request.
For consumers, the same, however, you'll want to make sure your offsets.topic.replication.factor (and transactions topic factor, if you use them) is higher than 1; otherwise, consumers will be unable to read anything (or transactions will not work, which are enabled by default in newer versions)
I've configured a cluster of Kafka brokers and a cluster of Zk instances using kafka_2.11-1.1.0 distribution archive.
For Kafka brokers I've configured config/server.properties
broker.id=1,2,3
zookeeper.connect=box1:2181,box2:2181,box3:2181
For Zk instances I've configured config/zookeeper.properties:
server.1=box1:2888:3888
server.2=box3:2888:3888
server.3=box3:2888:3888
I've created a basic producer and a basic consumer and I don't know why I am able to write messages / read messages even if I shut down all the Zookeeper
instances and have all the Kafka brokers up and running.
Even booting up new consumers, producers works without any issue.
I thought having a quorum of Zk instances is a vital point for a Kafka cluster.
For both consumer and producer, I've used following configuration:
bootrapServers=box1:9092,box2:9092,box3:9092
Thanks
I thought having a quorum of Zk instances is a vital point for a Kafka cluster.
Zookeeper quorum is vital for managing partition lists, leaders, etc. In general, ZK is necessary for management that is done by the cluster coordinator in the cluster.
Basically, right now (with ZK down), you cannot modify topics (as the partition metadata is stored in ZK), start up / shut down brokers (as they use ZK for discovery) and other similar operations.
Even booting up new consumers, producers works without any issue.
Producer/consumer operations reach out to brokers only. The broker instance can still append to the log, and can still communicate with other brokers to have replication. So it is possible to send a message, get it received by broker and saved to disk, with other brokers replicating (as they are continuously sending fetch requests to the leader (and they know who this partition's leader is because they saved that data when ZK was still running)).
In our production environment, we often see that the partitions go under-replicated while consuming the messages from the topics. We are using Kafka 0.11. From the documentation what is understand is
Configuration parameter replica.lag.max.messages was removed. Partition leaders will no longer consider the number of lagging messages when deciding which replicas are in sync.
Configuration parameter replica.lag.time.max.ms now refers not just to the time passed since last fetch request from the replica, but also to time since the replica last caught up. Replicas that are still fetching messages from leaders but did not catch up to the latest messages in replica.lag.time.max.ms will be considered out of sync.
How do we fix this issue? What are the different reasons for replicas go out of sync? In our scenario, we have all the Kafka brokers in the single RACK of the blade servers and all are using the same network with 10GBPS Ethernet(Simplex). I do not see any reason for the replicas to go out of sync due to the network.
We faced the same issue:
Solution was:
Restart the Zookeeper leader.
Restart the broker\brokers that are not replicating some of the partitions.
No data lose.
The issue is due to a faulty state in ZK, there was an opened issue on ZK for this, don't remember the number.
I faced the same issue on Kafka 2.0,
On restart Kafka controller node everything caught-up on the replicas.
But still looking for the reasons why few partitions are under-replicated whereas the other partitions on the same nodes for the same topic works good, and this issue i see on a random partitions.
Do NOT run reassignment for all topics together, consider running it for small portions.
Find the topic that has under-replicated partitions and where reassignment process can't be completed.
Set unclean.leader.election.enable to true for this topic.
Find under-replicated partition that stuck for this topic. Check its leader ID.
Stop the broker (just the service, not the instance).
Execute Preferred Replica Election (in yahoo/kafka-manager or manually).
Start the broker back.
Repeat for the rest of topics that have the same problem.
Also I tried this advice, it didn't help me: https://stackoverflow.com/a/51063607/1929406
I notice that when sending messages to kafka (a producer) the samples show connecting to port 9092 -- writing directly to a broker. When consuming the examples show connecting to port 2181, presumably zookeeper.
The latter makes sense--I want to read from "the cluster", letting zookeeper figure out which broker the client should communicate with, and managing such things as knowing who's alive/dead in the cluster.
Why wouldn't publish/writes work the same way, i.e. write to "the cluster" (via zookeeper)?
Am I understanding this correctly, that for producing I'm bypassing zookeeper (cluster knowledge) and must know producer nodes (and presumably figure out what to do if one fails)?
The "high level consumer" of Kafka uses Zookeeper to keep track of which partitions each member in a consumer group is consuming and sometimes to track which offsets were read in which partition. Since access to Zookeeper is required, we may as well use it to figure out where are the brokers...
In the new consumer (coming soon in the next release), Zookeeper is no longer needed, and consumers connect directly to brokers, just like producers currently do.
I have Ubuntu 14.04TS. I use Node.js->Kafka->Storm->MongoDB chain. With initial development, everything goes well. Messages are finally stored into mMngoDB.
In Kafka, I have one Zookeeper and broker0 in kakfa1. broker1 in kafka2. With Storm, Zookeeper, nimbus, and DRPC are located at storm1. Supervisor and worker are located at storm2.
Now the questions is when I do update storm1 and storm2. I stopped all processes of storm1 and storm2. I suppose Kafka will buffer the message from Node.js. After I restarted both storm1 and storm2, and redeployed topology, I found messages during storm1 storm2's, down and up, are lost. So indeed, Kafka does not keep persistence of messages during storm maintenance period.
In my mind, Kafka will remember the last index of the message it receive acknowledgement.
In all, how could I prevent message from lost when storm is under maintenance.