Re:Show HN: Pg_CRDT – an experimental CRDT extension for Pos...
by kiwicopple on Saturday December 10, 2022 @12:27PM
(#33934493)
> This could be as basic as a three column table (guid, doc, clock), the clock being a global Lamport
This would be achievable, but also perhaps quite strict - we're dictating a table structure rather than working with existing CRDT libs. That said, we could probably do this today, without any mods/extensions
> querying/indexing
Yeah, there is nothing in this implementation for that
> Without the ability to index properties, or do ad-hock queries, on the documents there is little reason to to the merging in-DB
One benefit is the ability to send "directly" to the database, but I agree that lack of additional features in the current form outweighs this benefit
> a serialised JSON
This is how we might do it with with "Realtime as an authority" approach (and is the most likely approach tbh).
Re:Show HN: Pg_CRDT – an experimental CRDT extension for Pos...
by aaronblohowiak on Saturday December 10, 2022 @12:16PM
(#33934389)
Read modify write outside the db seems worse though?
Without the ability to index properties, or do ad-hock queries, on the documents there is little reason to to the merging in-DB. It then makes more sense to merge outside the DB and save both the raw CRDT object and a serialised JSON next to it for querying. I assume that if you had found it's performance better, indexing+querying would have been the next step?
However I'm not surprised you found performance not ideal, I suspect a layer in-front of PG throttling writes is the only way forward for a real-time use case.
Personally I find the offline + eventual consistency the most interesting use case. The ability to sync a users document set to a local db (SQLite or a future mini-pg/supabase) with local querying and then merge them back later. This could be as basic as a three column table (guid, doc, clock), the clock being a global Lamport like clock to enable syncing only changed documents.