> This write amplification is large enough that our efforts to tune indexing throughput have started to hit diminishing returns.

> don't key on the ANN address. That is precisely the change turbopuffer v3 makes. As you can imagine, it is not a trivial change.

This is a direct parallel to how Postgres and Mysql built indexes.

Your design choice went from a Postgres design pattern to a Mysql one. The difference is the reindexing cost vs the lookup cost - Postgres optimized for lookup and Mysql does for indexing on writes. Or more accurately, Postgres was better with good schema design using joins & mysql was optimized for a bad design with less normalization where many indexes exist for the same table.

Postgres always points an index to a row-id within postgres which is an arbitrary value which changes on each update.

Mysql, always assuming the storage engine is pluggable, points to the primary index entry and adds an extra indirection to the lookup.

This means that you point the mysql index to a stable id, so unless you go update the primary key for a row, you won't have to update the indexes for all the attribute lookups you might have made to data.

I don't do databases any more that much, but the design for NIMBLE file format has a lot of quirks which are relevant to this specific idea (wide tables).

But the old Uber post about switching from Postgres to Mysql to prevent index amplification[1] is a direct mirror to this post.

[1] - https://www.uber.com/us/en/blog/postgres-to-mysql-migration/

Vector databases were always more about retrieval than either vectors or data storage. But the term stuck all too well and companies held on to it a tad too long. Sorry :)

AI has some of the craziest up and down cycles of tech I've ever seen

I’m developing a local “code graph mcp tool” (not yet published) and followed a similar path, though I may have been able to go further since I have fewer vectors in my database (even on projects with 50M LOC).

At first, I tried all those popular vector databases and was disappointed with their performance. In the end, the best and fastest solution turned out to be building a multi-database system on SQLite, compiled with everything related to multi-client operations removed. Only exclusive mode was left. Everything is as binary as possible. The index is completely separate — an IVF with pre-training — and is built on the GPU (250K vectors are built, processed, and saved in 4 seconds). Right now, my biggest problem is frequent data changes, and I need to implement optimizations to reduce recalculations.

So far, I haven’t seen any vector database implementations that are heading in the right direction. Maybe only Lancedb looks promising, but it’s too heavy for my needs.

I've really liked lancedb for similar use cases. Not just that it is OSS. But Lance treats ANN as a secondary index similar to what turbopuffer v3 does. Rows sit in fragments, and the vector index never moves them.

> The problem with a vector primary index

We've realized this a long time ago at TopK and built a flexible serverless search engine from scratch. Supports dense/sparse vectors, late interaction, lexical search, indexed regex, filtering, and custom scoring in one query.

- https://www.topk.io/blog/vector-dbs-are-the-wrong-abstractio... - https://www.topk.io/blog/topk-embed-v1

the multi-vector duplication thing makes sense, copying every attribute once per vector explodes quickly. what's the new primary index?

I'd want to see p99 at 1k+ QPS on the same scale

Waitint for the CEO of Qdrant to step in

Im not full read up on RAG pipelines, but has anyone ever tried to make the database a neural net itself? I.e get rid of any sort of traditional databases, and then you basically just have some sort of autoencoder?

I find very little reason to use a pure vector database for enterprise retrieval. We built an enterprise retrieval engine on top of a SQL database with native vector support, and the flexibility is something we cannot ignore. Vector similarity is just one query primitive alongside full text search, filters, joins, ordering and normal relational predicates. Tenant/app/collection isolation becomes part of the query itself. ACLs, document versions, categories, metadata constraints and temporal filters are ordinary predicates rather than something you have to bolt onto a vector store. SQL is already going to be part of almost any enterprise system. Adding a separate vector database introduces another moving part and syncing two system whenever you update your data is the most difficult thing to get right.

This sounds like the Postgres vs. InnoDB argument 10 years later. Postings pointed at physical location (the ANN slot), so every SPFresh rebalance rewrote every index touching that doc. InnoDB solved this by pointing secondary indexes at the PK and eating an extra lookup on read. Curious what that extra lookup costs you when it's an S3 GET instead of a B-tree hop.

"Updating one vector can move hundreds of attributes and their indexes" is basically Uber's 2016 Postgres write amplification post, but for search. Same fix too: stop pointing indexes at where the row lives.

So ANN becomes a secondary index that points at a doc ID, and vector search now needs a hop to complete. Do clusters keep their own copy of the vectors so the search itself stays local, and only result fetch pays the indirection? Otherwise cold p99 seems like it gets worse.

AI Slop. Will not read.

It would be nice to have a page that actually loads. This one doesn't. RIP.

UPDATE: It loads now, but it didn't when it was first posted. Traffic load on the server does matter.

Is it time to kill the database and replace it with a LLM optimized compiled version that simply implements the required API directly in (Rust) code, without any dynamic overhead? It probably will still be based of off a base design or a base file format.

Ultimately this system will encompass the whole OS, of course, but the DB might be the best place to start.