SendGrid engineer here.
IIRC, each of the three founders wrote components in the language that they were most productive in at the time. So there were services in Python, Perl, or PHP depending on who wrote it. Seems like a solid strategy to take the fastest path to MVP when you're still trying to prove there's a market for what you're building.
As the company grew, Python emerged as the most commonly chosen language for new projects.
As you said, the initial appeal of Go was its performance. An occasional side-effect of "fast" is needing less hardware to do the same work, so its efficiency was also attractive.
I think a key moment that got us interested in Go came at an internal hackathon. One of our engineers built a prototype replacement for a core piece of our MTA. It was something like 6x faster than the existing Python version. The ability to produce that in such a short period of time demonstrated that Go was also a surprisingly productive language.
These days, most of our engineering teams choose Go for new backend services -- even those that aren't performance-sensitive. However, we still have a lot of active development in Python (and Ruby).
Disclaimer: I've been at SendGrid about 4 years, so some of this is second-hand. I'll ask the founders if they're up for doing an article on our engineering blog about it.
I only use Python for the same reason I use Java: all the NLP / ML frameworks built by people who don't know Go yet.
This is yet another Python success story. You created a viable product with Python and moved critical parts to Go when there was real need for doing so.
Considering Go has only been your primary language for 9 months, as the beginner's style code base is optimized for all that Go has to offer you'll continue to see more performance gains, albeit probably not as significant as you saw between v1 and v2. Good on you for not being tied down with optimizations and actually releasing.
"With Python we often found ourselves delegating logic to the database layer purely for performance reasons. " -- you mean you wrote postgres functions.. or.. views!? gasp! heresy! Why would you migrate something working perfectly well in postgres away to Go, though? You probably didn't achieve any performance gains with that. Were postgres functions really gobbling that much memory up? Or.. maybe my assumptions are wrong here.