March 10, 2026
Benchmarks or bench-marketing?
MariaDB innovation: vector index performance
MariaDB says it beats Postgres at AI search—fans cheer, skeptics cry benchmarketing
TLDR: MariaDB 12.3 claims faster, more accurate AI-style vector search than both its 11.8 release and Postgres with pgvector, while using less CPU. Commenters split between applause for cheaper queries and accusations of biased, single‑session “benchmarketing,” demanding real‑world, multi‑user tests before declaring a winner.
MariaDB dropped fresh numbers claiming its new 12.3 release finds ‘similar stuff’ faster and more accurately than 11.8—and even edges Postgres with pgvector. Translation: AI-style search inside your database uses less CPU and gets better on bigger datasets. The MariaDB crowd arrived with victory gifs; Postgres die‑hards rolled their eyes.
Then the drama: the work was done by Small Datum and sponsored by the MariaDB Foundation, prompting cries of ‘benchmarketing.’ One top comment mocked the ‘fork of a fork of a fork’ of ann-benchmarks like it was open‑source family tree night. Others questioned a single‑session test on a 48‑core box with everything cached—‘nice for labs, not for production.’ The pgvector side nitpicked float16 settings; MariaDB fans pointed at the headline claim: lower CPU per query = smaller bills and fewer laptop jet engines.
In between the flames, pragmatists asked for reproducible runs, shared configs, and head‑to‑heads with real apps. Links flew to the MariaDB Foundation and the configs in the post, with calls to pit this against dedicated vector stores like FAISS and Milvus. But the vibe remained clear: Team MariaDB says ‘W secured,’ Team Postgres says ‘show us real‑world receipts.’
Key Points
- •MariaDB 12.3 shows improved recall-precision for vector search over MariaDB 11.8, especially on larger datasets (500k and 1,000k).
- •MariaDB 11.8 outperforms PostgreSQL 18.2 with pgvector 0.8.1 in recall-precision; MariaDB 12.3 is the top performer overall.
- •Lower CPU usage per query is cited as a key reason for MariaDB 12.3’s performance advantage.
- •Benchmarks used ann-benchmarks on dbpedia-openai-100k/500k/1000k-angular datasets, run in a single session with data fully cached.
- •Tests ran on a Hetzner ax162-s (48 cores, 128 GB RAM) with Ubuntu 22.04 and RAID 10 on two NVMe drives; all software was compiled from source.