Would ZooKeeper be a single point of failure?

If ZooKeeper cluster down, what’s the impact on data plane? Would the producing / consuming been blocked?

You can check our architecture page for details.

Basically, metadata store (can be ZooKeeper or etcd at the moment) is required component for a running cluster. If ZK ensemble doesn’t response for any reason, although you can still talk with Pulsar Broker, it can make little progress, because topic metadata should be stored in the metadata store.

Specifically, producing and consuming require accessing and manipulating metadata, so if ZK is unavailable, they will block.

In this point, you may think if ZK failed, the Pulsar cluster cannot serve normally. However, generally we don’t regard ZK cluster unavailable as a single point of failure. Because a ZK ensemble often consists by 3 nodes or 5 nodes, and it’s rare that the whole cluster is down. You may ask your operation team for its availability agreement.