This operator can be used to set up a highly available HDFS cluster. It implements three roles of the HDFS cluster:
Data Node - responsible for holding the actual data. IMPORTANT Currently the data is kept in
hostPathvolumes of 1GB and no configuration options are exposed to the user. Each data node has its own volume.
Journal Node - responsible for keeping track of HDFS blocks and used to perform failovers in case the active name node fails. For details see: https://hadoop.apache.org/docs/stable/hadoop-project-dist/hadoop-hdfs/HDFSHighAvailabilityWithQJM.html
Name Node - responsible for keeping track of HDFS blocks and providing access to the data.
The operator creates the following K8S objects per role group defined in the custom resource.
ClusterIP - used for intra-cluster communication.
ConfigMap - HDFS configuration files like
hdfs-site.xmlare defined here and mounted in the pods.
StatefulSet - where the replica count of each role group is defined. By default, a cluster will have 2 name nodes, 3 journal nodes and 3 data nodes.
In addition, a
NodePort service is created for each pod that exposes all container ports to the outside world (from the perspective of K8S).
In the custom resource you can specify the number of replicas per role group (name node, data node or journal node) but the operator will make sure that: * at least two name nodes are started * at least one journal node is started * no datanodes are started unless the number of replicas is greater than zero.
The cluster can be monitored with Prometheus from inside or outside the K8S cluster.
All services (with the exception of the Zookeeper daemon on the node names) run with the JMX exporter agent enabled and expose metrics on the
metrics port. This port is available form the container level up to the NodePort services.
The metrics endpoints are also used as liveliness probes by K8S.