True, horizontal scaling will always save you no matter how slow the front end is. Cost becomes significant at a certain level too though. For example, Google estimates each search query goes to over a thousand machines. If you need 100x 1000 machines to serve a query because the back end is PHP it adds up.
And you can make Python or even PHP fast if you try hard enough.
My argument is that the engineer overhead for Go and new breeds of Java frameworks are small enough that it makes no sense to use anything else if you're planning on scaling for real.
If you start with something else the cost of making a slow language fast and the multiples of extra machines you need costs far more than just using the faster language to start with
For the benchmark you posted, take a good look at the "realistic" vs "stripped" implementations and whether the test used an ORM. You'll quickly see that the realistic test applications with any kind of ORM are exclusively C#, Java, Go, and C++
And you can make Python or even PHP fast if you try hard enough.
My argument is that the engineer overhead for Go and new breeds of Java frameworks are small enough that it makes no sense to use anything else if you're planning on scaling for real.
If you start with something else the cost of making a slow language fast and the multiples of extra machines you need costs far more than just using the faster language to start with
For the benchmark you posted, take a good look at the "realistic" vs "stripped" implementations and whether the test used an ORM. You'll quickly see that the realistic test applications with any kind of ORM are exclusively C#, Java, Go, and C++