Recommended specs for on prem leaf nodes

Help
1

Continuing the discussion from Maximum lead node memory limit:

When we selected SingleStore (this was before my time, and I believe it was before the rebranding), the SingleStore engineer recommended a maximum of 4 CPUs and 64 GBs per leaf node. We would get better performance from many small leaf nodes rather than fewer, more robust, leaf nodes.

What are the current guidelines for node provisioning?

We only utilize 50% of our licenses with four-leaf nodes with four(4) (v)CPUs and 64 GB of RAM each. With over 8.4 TB of columnstore tables, most of which have 85% to 93% compression, which is utilized by four different internal (OLAP) and customer facing (usually hybrid OLTP/OLAP) applications. HA is enabled as well.

CPU is an almost constant 10% utilization while memory is almost always 60+% when the cluster is idle idle, even during idle periods. Considering we have less than 2 GBs of rowstore tables. We are running 7.5.12 and several of our commonly used (OLAP) tables are nearly 1 TB in size with 92% compression.

Thanks for any insights.

2

Hi @smaxwell,

There’s no recommended maximum size for a leaf node. We typically do better with larger node sizes, provided that you have enough I/O bandwidth per vCPU (per core).

E.g. if you have 1 SSDs per node, and 8 small nodes, that might be faster than if you had one large node with the same total cores and RAM as the 8 small nodes combined, but only 2 SSDs. You’d of course want to stripe your datadir file system over multiple SSDs or disks if you have multiple SSDs or disks, to get full I/O bandwidth.

Our typical users with larger-scale deployments are having good results with anything from 8 core and 32GB RAM leaf nodes to 56 core and 1TB RAM leaf nodes.

Many users on our managed service are in production with as little as 2 cores and 16 GB RAM, with a single leaf node. Sizing really depends on your workload, performance goals, and budget.