The document speaks of caching on Historicals and on Brokers and how these are kept in the heap. Easy enough, since it's the process heap. However, when doing streaming ingestion on Middle Managers, I'm a bit confused on where it logically resides. Does the realtime cache exist: 1. On the central process heap (one cache per node, requiring

cacheSize

space) 2. On the task heap (one cache per task, requiring

cacheSize * tasks

space) I suspect it's on the task heap, but I want to double check with someone more familiar before diving into the source code for a verdict.

You can enable caching on tasks, it uses memory from the heap of each task. See the

Peons can cache

section at the bottom of https://imply.io/developer/lessons/how-to-tune-apache-druid-for-speed-and-concurrency/ 🙂

The alternative (which I am using in production) is routing your realtime tasks (and only those) to an https://druid.apache.org/docs/latest/design/indexer - that's essentially a middle manager where all tasks are in-process. In my experience this scales much better for realtime, as you can share the query engine and cache.