Same reason as interpreters running system or performance sensitive things, you know, like servers and device drivers and codecs and the like. It's a swiss army knife when a fixed blade is all you needed. The large majority of simpler types of table data that you need to access in a server are tens if not hundreds of times faster implemented as a key value store for each table.

Just look at what people report when they moved from Python to Go - typically a ten times increase in performance.

SQL isn't quite as complex as python but it is interpreted and the latency is increased a great deal compared to a native key value store on an log system like used in rocks, leveldb and such.

If you don't care about wasting, incredible amounts, of resources, then SQL is great because it's so widely supported and known. If you want to use as little as possible AND provide minimal latency, there is nothing faster than a rocksdb on an SSD storage, that's what Facebook uses.