basically, if you want the best performance and your disks aren't super fast, you want to avoid hitting disk Druid uses columnar storage and memory mapping to minimize disk reads, so you just need to carve out enough free memory for that to work the amount you need depends on your working set— how much of your data do you expect to be involved in queries at any given slice of time?

the main things to think about there are how your data is partitioned (by time), clustered/secondary-partitioned (by some columns), and how many datasources (tables) you have

also how many columns you have & what proportion of columns are likely to be active for queries in any given slice of time

as you can see: fuzzy 🙂

generally, though, bottom line is 10% to 40% tends to work well for most people. although i have seen cases that require 100% and cases that are OK even on, like, 1%. so I'd advise keeping an open mind

it's always nice to deploy in a cloud environment so you can adjust to what hardware is best for your particular situaton