Run Mixed Workload using 2 Cassandra Instances on Each Node
This post has NOT been accepted by the mailing list yet.
We are trying to Decouple our Reporting DB from OLTP. Need urgent help on the feasibility of proposed solution for PRODUCTION.
Use Case: Currently, our OLTP and Reporting application and DB are same. Some CF are used for both OLTP and Reporting while others are solely used for Reporting.Every business transaction synchronously updates the main OLTP CF and asynchronously updates other Reporting CFs.
1. Decouple Reporting and OLTP such that Reporting load can't impact OLTP performance.
2. Scaling of Reporting and OLTP modules must be independent
3. OLTP client should not update all Reporting CFs. We generate Data Records on File sytem/shared disk.Reporting should use these Records to create Reporting DB.
4. Small customers may do OLTP and Reporting on same 3-node cluster. Bigger customers can be given an option to have dedicated OLTP and Reporting nodes. So, standard Hardware box should be usable for 3 deployments (OLTP,Reporting or OLTP+Reporting)
Note: Reporting is ad-hoc, may involve full table scans and does not involve Analytics. Data size is huge 2TB (OLTP+Reporting) per node.
Hardware : Standard deployment -3 node cluster with each node having 24 cores, 64GB RAM, 400GB * 6 SSDs in RAID5
1. Split OLTP and Reporting clients into two application components.
2. For small deployments where more than 3 nodes are not required:
A. Install 2 Cassandra instances on each node one for OLTP and other for Reporting
B. To distribute I/O load in 2:1 --Remove RAID5 (as Cassandra offers replication) and assign 4 disks as JBod for OLTP and 2 disks for Reporting
C. RAM is abundant and often under-utilized , so assign 8GB each for 2 Cassandra instance
D. To make sure that Reporting is not able to overload CPU, tune concurrent_reads,concurrent_writes
OLTP client will only write to OLTP DB and generate DB record. Reporting client will poll FS and populate Reporting DB in required format.
3. Larger customers can have Reporting clients and DB on dedicated physical nodes with all resources.
Is it ok to run 2 Cassandra instances on one node in Production system and limit CPU Usage,Disk I/O and RAM as suggested above?
Any other solution for above mentioned problem statement?