Question
The HBase bulkLoad task (a single table contains 26 TB data) has 210,000 maps and 10,000 reduce tasks, and the task fails.
Answer
ZooKeeper I/O bottleneck observation methods:
- On the monitoring page of Manager, check whether the number of ZooKeeper requests on a single node exceeds the upper limit.
- View ZooKeeper and HBase logs to check whether a large number of I/O Exception Timeout or SocketTimeout Exception exceptions occur.
Optimization suggestions:
- Change the number of ZooKeeper instances to 5 or more. You are advised to set peerType to observer to increase the number of observers.
- Control the number of concurrent maps of a single task or reduce the memory for running tasks on each node to lighten the node load.
- Upgrade ZooKeeper data disks, such as SSDs.