Hi, Marco!

Thanks for reaching out. The optimal way is to use an expensive Premium plan, and the other way is to have a predictive load on your app.

To solve a cold start problem, you should take into account application behavior, such as how many functions can run on one VM and how many VM`s deployed during particular spike operation(i.e. 10 requests/minute vs 1000 requests/min).
Find rush hours, understand conditions that predict spikes in usage, and deploy additional pinger() that will induct load that will result in 50 VM online before this spike.
You can ping me on twitter for more details.

Cheers.