This is a typical picture in htop of a starting ClickHouse:

One or two CPUs are saturated, but the rest sits idle. We 48T of data across ~240 tables on this machine and it takes ~18 minutes until ClickHouse is able to accept connections. We see around 35GB RSS at that point.

Not all stages are that bad, we also see this:

But it doesn't go anywhere near saturating either disks or CPU.

It feels like there's a room for more aggressive parallel startup.