I can explain it: write throughput has become
the bottleneck for an increasing number of applications and workloads. If you can't parse, process, index, and store real-time data at the rate it is created, or you can't load bulk data in a reasonable period of time, query performance doesn't matter. If you can only insert 100k records per second, how long will it take you to load a trillion records? A trillion row table fits on a single ordinary server these days and they aren't uncommon.
Note that even LSM-trees are only used for medium write intensity workloads. If you need maximum throughput for indexed access methods, newer database engines that use inductive indexing are on another level. I've seen a few implementations drive 10 million records per second through indexing and storage on an single EC2 instance (and with better query performance than LSM). These were not academic exercises either; they were invented to solve real-world problems LSM could not handle.
As for why anyone needs, say, 10 million inserts per second per server? There are quite a few modern workloads that generate many, many billions of new inserts per second if you run them at scale. Without the high throughput on a per server basis, managing the implied cluster size would become a serious challenge on its own. Server hardware can support very high ingest rates at very high storage densities if the software can drive it.
Note that even LSM-trees are only used for medium write intensity workloads. If you need maximum throughput for indexed access methods, newer database engines that use inductive indexing are on another level. I've seen a few implementations drive 10 million records per second through indexing and storage on an single EC2 instance (and with better query performance than LSM). These were not academic exercises either; they were invented to solve real-world problems LSM could not handle.
As for why anyone needs, say, 10 million inserts per second per server? There are quite a few modern workloads that generate many, many billions of new inserts per second if you run them at scale. Without the high throughput on a per server basis, managing the implied cluster size would become a serious challenge on its own. Server hardware can support very high ingest rates at very high storage densities if the software can drive it.