By 2014 when I left, we had a few petabytes of analytics data for a very small but high traffic set of customers. Could we query all of that at once within a reasonable online SLA? No. We partitioned and sharded the data easily and only queried the partitions we needed.
If I were to do this now and didn't need near real-time (what is real-time?) I'd use sqlite. Otherwise I'ld use trickle-n-flip on postgres or mysql. There are literally 10+ year-old books on this wrt RDBMS.
And yes, even with 2000 clients reaching billions of requests per day, only the top few stressed the system. The rest is long tail.