Er, are your indices currently spread over all the nodes data in your cluster?
And what did you mean specifically with "autoscaling this when the resource is being used by 90%"? What/which resource, and how are you measuring? Is your search IO or CPU or network bound, and if you know how did you find out? And if you don't know, that's first thing you need to find out!!
Also, where would you scale out to? Are you using cloud provider like Elastic Cloud, AWS, etc? Scaling costs more $$$, make sure you realise the economics. if you are using your own DC/hardware, more hardware also costs $$$.
In my personal experience the easiest, lowest hanging fruit is to do the analysis as suggested by Hemming.