Vertica in virtual enviroment - three-node cluster necessary

dzhao

We are presented an application/system by a vendor that utilize the Vertica database. We are told that we need to follow your recommendation to have a three-node cluster. However we have a Vmware enviroment that handles OS failure (e.g. CPU, RAM, ...etc) and we have NetAPP storage system that would handle Disk failure and throuput requirement. I have run those perf tests and the results are all good. In this case, do we still need three-node cluster, or will one-node cluster will be sufficient? I mean, even we still have three nodes they will be in the exactly same enviroment, competing with each other.

Of course that this one node will be much more powerful. I know that you do distributed processing, meaning that all nodes work at the same time. However will a VM with double the processing power be able to handle what three nodes can do (considering no waste of energy doing sync between the nodes)?

Our data is not heavy. Vendor estimated 2TB uncompressed data for 13-month

The vendor themselves don't have much experience with virtualization of your system so I figure I just ask experts here to have an authoritive answer.

Thanks,
Difan

Jim_Knicely

Hi,

Like you said, with a single node you'll lose all of the MPP features of Vertica. Also, you won't have any "High Availability" in the event the single VM failure. Other than those missteps, Vertica does run happily and really well on a single node .

See:
https://my.vertica.com/docs/9.1.x/HTML/index.htm#Authoring/InstallationGuide/BeforeYouInstall/platformhardwarereqts.htm

https://my.vertica.com/docs/9.1.x/HTML/index.htm#Authoring/ConceptsGuide/Other/High_Availability.htm

martinc123

is there a performance difference running vertica in VM (cloud), single node 12 cores , vs 3 nodes with 4 cores ? not counting high availability ?

Jim_Knicely

Hi,

From the docs:

The Vertica Analytics Platform is based on a massively parallel processing (MPP), shared-nothing architecture, in which the query processing workload is divided among all nodes of the Vertica database.

See:
https://my.vertica.com/docs/9.1.x/HTML/index.htm#Authoring/InstallationGuide/BeforeYouInstall/platformhardwarereqts.htm

So, in a single node you'd lose all of the MPP features available to Vertica! I would expect the biggest bottle neck on a single node would be the on the I/O controller and memory bus.

dzhao

Thanks Jim!