Major Description
- Scenario
Currently we have done a massive data loading. The input data is about 71GB in CSV format，and have about 88million records. When using carbondata, we do not use any dictionary encoding. Our testing environment has three nodes and each of them have 11 disks as yarn executor directory. We submit the loading command through JDBCServer.The JDBCServer instance have three executors in total, one on each node respectively. The loading takes about 10minutes (+-3min vary from each time).
We have observed the nmon information during the loading and find：
1. lots of CPU waits in the first half of loading;
2. only one single disk has many writes and almost reaches its bottleneck (Avg. 80M/s, Max. 150M/s on SAS Disk)
3. the other disks are quite idel
- Analyze
When do data loading, carbondata read and sort data locally(default scope) and write the temp files to local disk. In my case, there is only one executor in one node, so carbondata write all the temp file to one disk(container directory or yarn local directory), thus resulting into single disk hotspot.
- Modification
We should support multiple directory for writing temp files to avoid disk hotspot.
Ps: I have improved this in my environment and the result is pretty optimistic: the loading takes about 6minutes (10 minutes before improving).
