Consider the below scenario:
I have a Kakfa broker cluster(localhost:9002,localhost:9003,localhost:9004,localhost:9005).
Let's say localhost:9002 is my primary(leader) for the cluster.
Now my producer is producing data and sending it to the broker(localhost:9002).
If my primary broker(localhost:9002) goes down, with the help of Zookeeper or some other consensus algorithm new leader will be elected(consider localhost:9003 is now the new leader).
So, in the above scenario can someone please explain to me how the Kafka client(producer) will get notified about the new broker configuration(localhost:9003) and how it will connect to the new leaders and start producing data again.
Kafka clients are receiving the necessary meta information from the cluster automatically on each request when reading from or writing to a topic in case of a leadership change.
In general, the client sends a (read/write) request to one of the bootstrap server, listed in the configuration bootstrap.servers. This initial request (hence called bootstrap) returns the details on which broker the topic partition leader is located so that the client can communicate directly with that broker. Each individual broker contains all meta information for the entire cluster, meaning also having the knowledge on the partition leader of other brokers.
Now, if one of your broker goes down and the leadership of a topic partition switches, your producer will get notified about it through that mechanism.
There is a KafkaProducer configuration called metadata.max.age.ms which you can modify to update metadata on your producer even if there is no leadership change happening:
"Controls how long the producer will cache metadata for a topic that's idle. If the elapsed time since a topic was last produced to exceeds the metadata idle duration, then the topic's metadata is forgotten and the next access to it will force a metadata fetch request."
Just a few notes on your question:
The term "Kafka broker cluster" does not really exists. You have a Kafka cluster containing one or multiple Kafka brokers.
You do not have a broker as a "primary(leader) for the cluster" but you have for each TopicPartition a leader. Maybe you mean the Controller which is located on one of the brokers within your cluster.
Related
I am seeing org.apache.kafka.common.errors.NotLeaderForPartitionException on my producer which I understand happens when producer tries to produce messages to a broker which is not a leader for the partition.
Does that mean each time a leader fulfills a write request it first checks if its the leader or not?
If yes does that translates to a zookeeper request for every write request to know if the node is the leader?
How Producer Get MetaData About Brokers
The producer sends a meta request with a list of topics to one of the brokers you supplied when configuring the producer.
The response from the broker contains a list of partitions in those topics and the leader for each partition. The producer caches this information and therefore, it knows where to redirect the messages.
When Producer Will Refresh MetaData
I think this depends what kafka client you used.There are some small differents between ruby, java or other kafka client.for example, in java:
producer will fetch metadata when client initialize,then period update it depends on expiration time.
producer also will force update metadata when request error occured,such as InvalidMetadataException.
But in ruby-kafka client, it usually refresh metadata when error occured or initialize.
I understand producer/consumers need to talk to brokers to know leader for partition. Brokers talk to zk to tell they joined the cluster.
Is it true that
Brokers know who is the leader for a given partition from zk
zk detects broker left/died. Then it re-elects leader and sends new leader info to all brokers
Question:
why do we need brokers to communicate with each other? Is it just
so tehy can move partitions around or do they also query metadata from each other. If so what would be example of metadata exchange
Producers/ consumers request metadata from one of the brokers ( as each one of them caches it) and that is how they know who is the leader for a partition.
Regarding "is it true that" section:
Brokers know who is the leader for the given partition thanks to zk and one of them. To be more precise, one of them decides who will be a leader. That broker is called controller. The first broker that connects to zookeeper becomes a controller and his role is to decide which broker will be a leader and which ones will be replicas and to inform them about it. Controller itself is not excluded from this process. It is a broker like any other with this special responsibilities of choosing leaders and replicas
zk indeed detects when a broker dies/ leaves but it doesn't reelect leader. It is controller responsibility. When one of the brokers leaves a cluster, controller gets information from zk and it starts reassignment
About your question - brokers do communicate with each other ( replicas are reading the messages from leaders, controller is informing other brokers about changes), but they do not exchange metadata among themselves - they write metadata to a zookeeper
A Broker is a Kafka server that runs in a Kafka Cluster
"A Kafka cluster is made up of multiple Kafka Brokers. Each Kafka Broker has a unique ID (number). Kafka Brokers contain topic log partitions. Connecting to one broker bootstraps a client to the entire Kafka cluster"
Each broker holds a number of partitions and each of these partitions can be either a leader or a replica for a topic. All writes and reads to a topic go through the leader and the leader coordinates updating replicas with new data. If a leader fails, a replica takes over as the new leader.
Does Kafka broker store metadata which producer API uses (e.g. which partitions are leader for a topic etc.)? As per my understanding this metadata is stored in Zookeeper , is it correct? If it is true then how Brokers are updated by Zookeeper with latest information?
All Kafka brokers can answer a metadata request that describes the current state of the cluster: what topics there are, which partitions those topics have, which broker is the leader for those partitions etc.
ZooKeeper is responsible for:
Electing a controller broker - and making sure there is only one
Cluster membership - allowing brokers to join a cluster
Topic configuration - which topics exist, how many partitions each has, where are the replicas, who is the preferred leader, what configuration overrides are set for each topic
Quotas - how much data is each client allowed to read and write
ACLs - who is allowed to read and write to which topic
There is regular communication between Kafka and ZooKeeper such that ZooKeeper knows a Kafka broker is still alive (ZooKeeper heartbeat mechanism) and also in response to events such as a topic being created or a replica falling out of sync for a topic-partition.
Kafka is a distributed system and is built to use Zookeeper which is responsible for controller election, topic configuration, clustering etc.
More precisely, Zookeeper initiates controller election. The controller broker is a single broker in the Kafka cluster which takes care of leader broker and followers for every partition. When a particular broker is taken down, the controller lets other replicas know (in order to handle partition leaders etc). Moreover, when the controller fails then Zookeeper initiates new elections in order to elect the new broker which will act as the controller.
Furthermore, Zookeeper knows which brokers are part of the Kafka cluster and which are still alive. Similarly, it is also aware of topic-specific information such as which topics exist, how many partitions each has, where are the replicas and so on.
Zookeeper also stores information regarding quotas and ACLs, i.e. what volume of data each client is allowed to consume/produce and also, who is allowed to consume or produce from a particular topic.
The producers send messages by setting up a list of Kafka Broker as follows.
props.put("bootstrap.servers", "127.0.0.1:9092,127.0.0.1:9092,127.0.0.1:9092");
I wonder "producers" how to know that which of the three brokers knew which one had a partition leader.
For a typical distributed server, either you have a load bearing server or have a virtual IP, but for Kafka, how is it loaded?
Does the producers program try to connect to one broker at random and look for a broker with a partition leader?
A Kafka cluster contains multiple broker instances. At any given time, exactly one broker is the leader while the remaining are the in-sync-replicas (ISR) which contain the replicated data. When the leader broker is taken down unexpectedly, one of the ISR becomes the leader.
Kafka chooses one broker’s partition’s replicas as leader using ZooKeeper. When a producer publishes a message to a partition in a topic, it is forwarded to its leader.
According to Kafka documentation:
The partitions of the log are distributed over the servers in the
Kafka cluster with each server handling data and requests for a share
of the partitions. Each partition is replicated across a configurable
number of servers for fault tolerance.
Each partition has one server which acts as the "leader" and zero or
more servers which act as "followers". The leader handles all read and
write requests for the partition while the followers passively
replicate the leader. If the leader fails, one of the followers will
automatically become the new leader. Each server acts as a leader for
some of its partitions and a follower for others so load is well
balanced within the cluster.
You can find topic and partition leader using this piece of code.
EDIT:
The producer sends a meta request with a list of topics to one of the brokers you supplied when configuring the producer.
The response from the broker contains a list of partitions in those topics and the leader for each partition. The producer caches this information and therefore, it knows where to redirect the messages.
It's quite an old question but I have the same question and after researched, I want to share the answer cuz I hope it can help others.
To determine leader of a partition, producer uses a request type called a metadata request, which includes a list of topics the producer is interested in.
The broker will response specifies which partitions exist in the topics, the replicas for each partition, and which replica is the leader.
Metadata requests can be sent to any broker because all brokers have a metadata cache that contains this information.
I have a set of Kafka broker instances running as a cluster. I have a client that is producing data to Kafka:
props.put("metadata.broker.list", "broker1:9092,broker2:9092,broker3:9092");
When we monitor using tcpdump, I can see that only the connections to broker1 and broker2 are ESTABLISHED while for the broker3, there is no connection from my producer. I have a single topic with just one partition.
My questions:
How is the relation between number of brokers and topic partitions? Should I always have number of brokers = number of partitons?
Why in my case, I'm not able to connect to broker3? or atleast my network monitoring does not show that a connection from my Producer is established with broker3?
It would be great if I could get some deeper insight into how the connection to the brokers work from a Producer stand point.
Obviously, your producer does not need to connect to broker3 :)
I'll try to explain you what happens when you are producing data to Kafka:
You spin up some brokers, let's say 3, then create some topic foo with 2 partitions, replication factor 2. Quite simple example, yet could be a real case for someone.
You create a producer with metadata.broker.list (or bootstrap.servers in new producer) configured to these brokers. Worth mentioning, you don't necessarily have to specify all the brokers in your cluster, in fact you can specify only 1 of them and it will still work. I'll explain this in a bit too.
You send a message to topic foo using your producer.
The producer looks up its local metadata cache to see what brokers are leaders for each partition of topic foo and how many partitions does your foo topic have. As this is the first send to the producer, local cache contains nothing.
Producer sends a TopicMetadataRequest to each broker in metadata.broker.list sequentially until first successful response. That's why I mentioned 1 broker in that list would work as long as it's alive.
Returned TopicMetadataResponse will contain the information about requested topics, in your case it's foo and brokers in the cluster. Basically, this response contains the following:
list of brokers in the cluster, where each broker has an ID, host and port. This list may not contain the entire list of brokers in the cluster, but should contain at least the list of brokers that are responsible for servicing the subject topic.
list of topic metadata, where each entry has topic name, number of partitions, leader broker ID for each partition and ISR broker IDs for each partition.
Based on TopicMetadataResponse your producer builds up its local cache and now knows exactly that the request for topic foo partition 0 should go to broker X.
Based on number of partitions in a topic, producer partitions your message and accumulates it with the knowledge that it should be sent as a part of batch to some broker.
When the batch is full or linger.ms timeout passes, your producer flushes the batch to the broker. By "flushes" I mean "opens a new connection to a broker or reuses an existing one, and sends the ProduceRequest".
The producer does not need to open unnecessary connections to all brokers, as the topic you are producing to may not be serviced by some brokers, and your cluster could be quite large. Imagine a 1000 broker cluster with lots of topics, but one of topics has just one partition - you only need that one connection, not 1000.
In your particular case I'm not 100% sure why you have 2 open connections to brokers, if you have just a single partition, but I assume one connection was opened during metadata discovery and was cached for reusing, and the second one is the actual broker connection to produce data. However, I might be wrong in this case.
But anyway, there is no need at all to have a connection for the third broker.
Regarding your question about "Should I always have number of brokers = number of partitons?" the answer is most likely no. If you explain what you are trying to achieve, maybe I'll be able to point you to the right direction, but this is too broad to explain in general. I recommend reading this to clarify things.
UPD to answer the question in comment:
Metadata cache is updated in 2 cases:
If producer fails to communicate with broker for any reason - this includes the case when the broker is not reachable at all and when broker responds with an error (like "I'm not leader for this partition anymore, go away")
If no failures happen, the client still refreshes metadata every metadata.max.age.ms (https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/clients/CommonClientConfigs.java#L42-L43) to discover new brokers and partitions itself.