Kafka Broker元数据存储疑问及Zookeeper更新机制咨询

Great questions! Let's break this down clearly and concisely:

1. Does a Kafka Broker store metadata? #

Yes, Kafka Brokers do store metadata—but we need to distinguish between local cached metadata and the authoritative source:

• Each broker maintains a local cache of cluster-wide metadata, including topic partition layouts, leader/follower assignments for all partitions, and the list of active brokers in the cluster.
• Additionally, a broker stores metadata specific to the partitions it hosts, such as the current state of its replicas (leader, in-sync follower, etc.) and log segment details.

2. Is the metadata used by the Producer API (e.g., which partitions are leaders) stored in Kafka Brokers? And is your understanding that it's stored in ZooKeeper correct? #

You're absolutely right that the authoritative source of this core metadata lives in ZooKeeper (note: newer Kafka versions support KRaft mode without ZooKeeper, but we'll focus on the ZooKeeper-based setup here).

Here's the full flow for producers:

• Producers never connect directly to ZooKeeper. Instead, they request metadata from a Kafka Broker (typically the bootstrap broker configured in their settings).
• The broker fetches and caches the latest metadata from ZooKeeper, then serves this cached data to the producer.
• So while producers get the metadata from a broker, the single source of truth for that data is always ZooKeeper.

3. How does ZooKeeper push the latest information to Brokers? #

ZooKeeper relies on a Watcher mechanism to notify brokers of metadata changes, and here's how it works:

• When a broker starts up, it registers Watchers on specific ZooKeeper nodes that hold critical cluster metadata—for example:
  • Nodes representing topics and their partition configurations
  • Nodes tracking the list of active brokers in the cluster
  • Nodes storing leader/follower assignments for partitions
• When any change occurs to these nodes (e.g., a new topic is created, a leader fails over, a broker joins or leaves the cluster), ZooKeeper triggers the corresponding Watcher and sends an immediate notification to the broker.
• Upon receiving the notification, the broker initiates a full fetch of the updated metadata from ZooKeeper, updates its local cache, and then makes this fresh metadata available to producers, consumers, and other brokers in the cluster.

内容的提问来源于stack exchange，提问作者Aman