Hi Puneet,
Most Your understanding is correct. And from our testing, SQL will be easily being the bottleneck. A few clarification:
1. Currently 3 Headnodes only for simplicity, we will support more in future
2. Other two headnodes actually are doing things as different services may active on different headnodes, it includes services:
- Scheduler Service: scheduling jobs
- Monitoring Service: receive metric data from all compute nodes and do aggregation
- Diagnostics Servcie
- Stateful management service: Do management operations
- Session Service: serve SOA job
And meanwhile there are stateless service that running on all headnodes and being active all the time, it includes services:
- Stateless Management Service: serve client request and client triggered operations such as node offline/online ...
- Scheduler REST service: serve client job request
- Naming Service
As we have disabled loading balancing, the service only failover when it encounters error. THus you are able to mannually distribute primary instance of a service on different headnode for example: Primary Scheduler service on one headnode, other primary
services on the other two to have better load balance between the three headnodes
Qiufang Shi
