Comment by Gluber | Hacker News Reader

Comment by Gluber | Hacker News Reader

I tend to agree with quite a few points in the article, but some topics warrant some careful scrutiny.

* As a message queue: Only if your required features are very basic, like if you need cluster communication and run your own coordination protocol on top.

* High Volume Time Series: TimeScale works, but composes badly with other workloads on the same DB server ( from an operational perspective at scale )

* Vector Database: The same issues as with TimeScale.. PgVector for example lives in its own seperate "world" and the query planner sees it as a very opaque thing. Forget about adding vector storage to an existing high volume db, that must server other complex queries.. PGVector will either trash your caches, or take over your cpu so that workloads that used to work fine stall. This is IMO not a pgvector problem itself ( Kudos to those guys ) but rather that postgresql extension apis are not very good at exposing custom costs and tradeoffs to the system as a whole.

* Raw Data: Works for small files... why anyone would want to store large amounts of data in it would be a mystery, where it shines is accessing LOTS of small files where internal caching etc help a lot compared to raw filesystem access ( also a bit dependent on the filesystem and its tuning though )

* Microservice: If your service is ONLY exposing json data from some database model, then it should not exist at all IMO. Create a view and be done with it.

Also to note: (Not a fault of PGVector again just a limit of our algorithmic knowledge) PGVector does HSNW or IVFlat indices ... (there is nothing better persistent) however it breaks down with high latency at LARGE amounts of vectors ( 100MIO+ ) that seems like a high ceiling, but when designing production RAG systems, you tend to do per chunk embeddings, or even visual patch embeddings... e.g one page of a document becomes 1024 vectors in itself (for visual patch embeddings ) ... so you hit those limits at 100000 pages already.. something larger organizations definitly have.

I would keep a per-document summary, then dive down only into the filtered set. This is more production grade than selecting from billions of chunks.

> Microservice: If your service is ONLY exposing json data from some database model, then it should not exist at all IMO. Create a view and be done with it.

Yeah, this has nothing to do with Postgres. If the service is accessing a database that isn't internal to the service, then that database is already a standalone service in itself.

I like to consider Postgres the starting point for all of these things, that can be outgrown and replaced when appropriate. I do love just shoving everything in Postgres and seeing that I only end up needing a few additional dedicated services as the product groups. Redis is usually the next pickup for me.

Sure, thats a good way of working.. I just have the experience when handing over a project ( consulting ) anything i have put in place will never get replaced or kept for too long outgrowing its capacity by far, and offset with huge expenses in hardware or operations. Technically not my problem anymore ( except when it breaks on a maintenance contract ) but i still like to avoid it early if i can

Reply on news.ycombinator.com