Epic: Pageserver archived timeline storage optimization

[arpad-m]

Motivation

We have timeline offloading now thanks to #8088, thanks to which offloaded timelines don't incur much local nvme costs, memory costs or background CPU any more. However, they still cause costs on remote storage. We don't do any special compression on them, and we don't run gc on offloaded timelines either.

The RFC has introduced the concept of flattening of a timeline, to make it be efficient in storage.

Tasks

Tasks

Preview Give feedback

1. implement timeline flattening (including some way to block offload for it)
2. try to run flattening right before offload during compaction. only once no flattening tasks are left, do the offload.
3. implement heuristics to unoffload timeline and run flatten/gc on it if it falls out of pitr etc
4. timeline warmup API
5. tests: after warming up, a read workload should not result in any on-demand downloads

Other related tasks and Epics

• Earlier epic: Epic: Pageserver Timeline Archival #8088
• RFC: rfcs: add RFC for timeline archival #8221