You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Oct 30, 2021. It is now read-only.
Transferring braindump notes to here. This issue affects both carmen-cache and carmen, but the problem originates mostly from the grid cache outward so capturing here for now.
Why the problem exists at????!!!
Feature storage is sharded right now not just for historical reasons but because the grid storage in our PBF format currently limits us 20-bit feature IDs.
Even with a larger feature ID space (e.g. if we went from 53 bits => 64 bits) we would probably encounter feature ID collisions at which point we'd need to store them in a nested fashion with some kind of secondary factor (like cover zxy) to disambiguate where possible.
Example:
feature A 120951912 \
+----> store together, distinguish by id + zxy
feature B 120951912 /
Rough braindump of how to untangle
We'd want unlimited feature ID storage space
We'd want enforcement of feature ID uniqueness (per index) at indexing time
We'd want to undo "sharded" feature storage
We'd want to stop using uint-64 as a way to transfer cover data into and out of carmen-cache -- we'd probably define a new PBF message format (variable length, allowing for any size feature IDs) for encoding this information
I think these would all be very positive changes, just a pretty solid, invasive refactor