ZooKeeper

Apache Druid uses Apache ZooKeeper (ZK) for management of current cluster state.

Minimum ZooKeeper versions

Apache Druid supports ZooKeeper versions 3.5.x and above.

Note: Starting with Apache Druid 0.22.0, support for ZooKeeper 3.4.x has been removed

ZooKeeper Operations

The operations that happen over ZK are

  1. Coordinator leader election
  2. Segment “publishing” protocol from Historical
  3. Segment load/drop protocol between Coordinator and Historical
  4. Overlord leader election
  5. Overlord and MiddleManager task management

Coordinator Leader Election

We use the Curator LeaderLatch recipe to perform leader election at path

  1. ${druid.zk.paths.coordinatorPath}/_COORDINATOR

Segment “publishing” protocol from Historical and Realtime

The announcementsPath and servedSegmentsPath are used for this.

All Historical processes publish themselves on the announcementsPath, specifically, they will create an ephemeral znode at

  1. ${druid.zk.paths.announcementsPath}/${druid.host}

Which signifies that they exist. They will also subsequently create a permanent znode at

  1. ${druid.zk.paths.servedSegmentsPath}/${druid.host}

And as they load up segments, they will attach ephemeral znodes that look like

  1. ${druid.zk.paths.servedSegmentsPath}/${druid.host}/_segment_identifier_

Processes like the Coordinator and Broker can then watch these paths to see which processes are currently serving which segments.

Segment load/drop protocol between Coordinator and Historical

The loadQueuePath is used for this.

When the Coordinator decides that a Historical process should load or drop a segment, it writes an ephemeral znode to

  1. ${druid.zk.paths.loadQueuePath}/_host_of_historical_process/_segment_identifier

This znode will contain a payload that indicates to the Historical process what it should do with the given segment. When the Historical process is done with the work, it will delete the znode in order to signify to the Coordinator that it is complete.