design: Partially tiered datastructures

[dranikpg]

Assumptions:

1. HINCRBY[FLOAT] are the only read-write commands. As they work on small values (integers/floats), we can assume that all commands are either: exclusively read, exclusively write or delete.

Definition. Very similar to tiered values, somewhat simpler:

1. A datastructure can have multiple nodes that are offloadable and are identified by a unique uint64_t
2. A node can be:
   • In memory with optionally a pending disk write
   • On disk with optionally a series of pending disk reads
3. The mechanics of reading/writing/deleting are the same as with tiered values
4. Uploading/offloading is now also handled by the decoder/datastucture

Identifier problems:

1. Reads can finish after a value has been aready overwritten in memory, so uploading is no longer possible
   -> Reusable identifiers (including allocation addresses) can only be used with some sort of versioning.
   The versioning can be performed by OpManager by marking read-chains as stale

[romange]

Reads can finish after a value has been aready overwritten in memory,

Don't we "delete" entries from tiered storage on writes? Which in turn signals the tiered component to avoid uploading pending reads for values that are deleted?