You shouldn't completely ignore performance all the way through and assume you can sprinkle it on top later, but you shouldn't spend months squeezing the last few ms out of a feature that the client will change / remove when they try it anyway.
> Really, you are gonna refactor a 100k Java codebase to be fast when its done?
Maybe? Depends where the problems are. If the problems are huge architectural ones, no that sounds terrible. If the problem can be solved by later improving an inner loop somewhere and getting things solved then sure.
People used Instagram at first because it was fast.
And they were quickly bought out for a billion.
Nobody would have used snapchat if it didn't take photos but the site loaded in 1ms. Nobody would use it if it had loads more features and took 2 hours.
Speed is a part of it, and must be traded off with all the other parts of the user experience.
I'm currently leading a team whose primary product is the front end service for the core SaaS business that does billions of monthly impressions (hundreds of millions of uniques). This product is over 10 years old, written in Spring, and has around 400,000 lines of Java.
A couple years ago, we underwent a massive overhaul to the system and refactored the application to use our realtime steaming platform instead of using MySQL+Solr.
It took some time to get it done, mainly because at the time there were about 8 years of features baked into a monolithic web service.
Ultimately, we were successful in the endeavor. It took about 2.5 years and our response times are dramatically faster (and less expensive in terms of infrastructure cost).
So yeah, you can do it. It's expensive and takes time if you wait too long. And you need full support from the business. And some companies don't survive it. YMMV.