Availability Domain

  • 27 October 2015
  • 1 reply
  • 5935 views

Badge +4
Hi all,

regarding the chapter dedicated to availability domain in Nutanix Bible (http://nutanixbible.com/index.html#availability-domains-WnIbC8TM) I have questions from a customer i was not able to answer.
This chapter begins by : Availability Domains (aka node/block/rack awareness) is a key struct for distributed systems to abide by for determining component and data placement.

Node awareness is the default feature : ok
I understand block awareness as described in the bible.
But nowhere else bible's team talk about rack awareness...

I will be audacious : would it be possible to create a Nutanix cluster on two sites (latency = 60microsecondes) with rack awareness (rack = site) ?

and more audacious : a customer with three sites (latency of each with the others = 60microsecondes). Is that to create a 3 sites cluster with rack awareness ? In case of failure of a site, the reconstruction will be done between the 2 stayed alive.

Is that possible ?
if yes how ?
if not why ?

regards
Herve

1 reply

Userlevel 2
Badge +12
Hi  --

Today Node / Block awareness is built into the software. Rack awareness is achieved by striping the cluster across racks and leveraging block awareness by which each "block" is actually in a different rack. This will be natively integrated in the future, however in most cases it is best practice to have a cluster within a rack to keep connectivity to 1 hop between nodes.

To answer your question... The brief answer is that while "possible" you don't want to do it (nor would it likely be supported).

It really comes down to quorum. In the two site scenario, if the network goes down between the sites (e.g. 8 nodes site A, 8 nodes site 😎, only the side with quorum (e.g majory if Zookeeper nodes) would remain up. The other side would be down since it wouldn't have a majority (we require quorum to eliminate any split brain scenarios).

With three sites, it's possible that a connectivity outage between all sites would cause all sites to go down (e.g. each site has 1 zookeeper node meaning none would keep quorum). If you just lost a site (no network outage), the other two sites would remain up.

So a long winded answer, but all in all you don't want to do it. If you want to survive a site failure with 0 RPO I'd go with Metro and EC (to minimize the overhead of dual RF)

Hope that helps and let me know if you have any other questions!

--Steven

Reply