December 30, 2025
Postgres vs Pinecone: Fight Night
Postgres extension complements pgvector for performance and scale
Postgres gets a turbo boost for AI search; readers cheer ditching Pinecone, grumble about AWS locks
TLDR: Timescale’s pgvectorscale supercharges Postgres for AI-style “find similar” searches, posting huge speed and cost wins over Pinecone in a benchmark. Commenters celebrate sticking with Postgres and hybrid search, while skeptics gripe about AWS RDS support and vendor tests, asking if managed services will catch up.
Timescale dropped pgvectorscale, a new add‑on that supercharges Postgres for AI‑style “find similar things” searches. Think: faster lookups, cheaper storage, and smarter filtering by labels (aka tags). It borrows ideas from Microsoft’s DiskANN research, adds a new index, squeezes data with a fresh compression trick, and — the headline claim — beats Pinecone’s popular index in a big benchmark: 28x lower latency, 16x more throughput, and 75% less cost when self‑hosting on AWS. It’s written in Rust, not C, which some folks cheer as “modern vibes.” Also noted: Intel Mac builds are a no‑go for now, and that raised a few eyebrows.
The crowd reaction? One user says this was the moment they stuck with Postgres and skipped “yet another database.” Others post receipts with a head‑to‑head blog link and a performance thread link. The drama: a cloud‑side reality check — “Is this still unsupported in RDS?” — sparking jokes that RDS stands for “Really Doesn’t Support.” Meanwhile, Timescale points to their text search add‑on link, hyping a hybrid search one‑stop shop. Memes flew about “DiskANN” sounding like a boy band, ARM‑Mac owners flexed, and skeptics questioned vendor benchmarks. The mood? Excited, a little spicy, and very Postgres‑pilled.
Key Points
- •pgvectorscale is a Rust-based PostgreSQL extension that complements pgvector to improve performance and reduce storage costs for AI embedding search.
- •It introduces StreamingDiskANN, Statistical Binary Quantization, and label-based filtered vector search, drawing on Microsoft research and Timescale innovations.
- •Benchmarks on 50M Cohere embeddings (768 dims) show 28x lower p95 latency and 16x higher throughput versus Pinecone’s s1 index at 99% recall.
- •Self-hosting PostgreSQL with pgvector+pgvectorscale on AWS EC2 reportedly costs 75% less than Pinecone in the benchmark context.
- •Installation options include a pre-built Docker image or building from source; macOS x86 builds are currently unsupported, with ARM Mac, Linux, or Docker as alternatives.