Skip to main content

Buyer guide

pgvector Production Playbook

A checklist for running pgvector in production: the host's release, index and vector types, filtered query plans and pinned upgrades.

Reading time
2 min read
Sources
9 cited
In this piece
pgvector−29 Negative

Confirm the pgvector release your host ships, choose index and vector types for your memory budget, test filtered queries against the planner at production size, and pin releases before upgrades. Each step cites pgvector's README and CHANGELOG, the hosts' own pages and public accounts from teams running it.

When you need this

  • A RAG or search feature is moving to production on Postgres.
  • Filtered vector queries have slowed as data grew.
  • The team is choosing a managed Postgres host for pgvector.

Evaluation criteria

  1. Host release

    Which pgvector release does your host ship, and how fast does it update?

    Look forCloud SQL lists 0.8.5 on PostgreSQL 13 and later. Azure needs the extension on the allowlist, and one team on Azure was held on 0.8.2 with an HNSW vacuum bug fixed in 0.8.3.

  2. Index type and memory

    Does the index build fit your memory?

    Look forHNSW builds slower and uses more memory than IVFFlat, and builds faster when the graph fits in maintenance_work_mem. One practitioner reports 10+ GB of RAM for an HNSW build on a few million vectors.

  3. Vector type

    Can you store at half precision?

    Look forhalfvec takes 2 * dimensions + 8 bytes against 4 * dimensions + 8 for vector. One benchmark's HNSW index fell from 38 GB to 19 GB with halfvec, and another developer saw no change in results.

  4. Filtered plans

    Does the planner use your index when filters apply?

    Look forRun EXPLAIN on production-sized data. Enable iterative index scans from 0.8.0 and use partial indexes for low-cardinality filters. Mem0's indexes recorded 0 scans under tenant filters.

  5. Upgrades

    Are images and releases pinned?

    Look forPin the Docker tag: a floating pg17 tag moved from 0.8.2 to 0.8.4 and broke one team's CI index build. Read the CHANGELOG before each upgrade.

Questions to ask vendors

  1. Which pgvector release does our host run today?
  2. Do our filtered queries use the index at production size?
  3. How much memory does an index rebuild need on our largest table?
  4. Who pins and upgrades the pgvector image?

Red flags

  • Filtered queries never checked with EXPLAIN.
  • A floating Docker tag in production or CI.
  • maintenance_work_mem left at the default for large index builds.
  • A managed host two releases behind on HNSW fixes.

Products in this piece

Sources

Official documentation
4
Public source
5
  1. pgvector on GitHub: READMEOfficial documentationraw.githubusercontent.com
  2. pgvector on GitHub: CHANGELOGOfficial documentationraw.githubusercontent.com
  3. Google Cloud: Cloud SQL for PostgreSQL extensionsOfficial documentationcloud.google.com
  4. Microsoft Learn: pgvector on Azure Database for PostgreSQL flexible serverOfficial documentationlearn.microsoft.com
  5. Mem0 blog: How we cut vector search latency by 70x (Jul 30, 2026)Public sourcemem0.ai
  6. digdir/genki issue #186: pgvector 0.8.2 on Azure and HNSW vacuum (Oct 29, 2026)Public sourcegithub.com
  7. pgvector issue #995: index build memory after 0.8.4 (Jul 1, 2026)Public sourcegithub.com
  8. pgvector issue #967: halfvec index size and compiler versions (Mar 9, 2026)Public sourcegithub.com
  9. pgvector issue #916: planner choices under a 1% filter (Oct 26, 2025)Public sourcegithub.com

Keep reading

  1. Researchpgvector in Production: Where Postgres Holds Up and Where Teams Move OutPublic accounts from Discourse, Instacart and Clarvo against those from Mem0 and Halcyon, and what decides the difference: filters, the query planner and scale.
  2. Comparisonpgvector vs Pinecone: Vectors in Your Postgres or a Managed Servicepgvector keeps vectors in the Postgres you run with no licence fee, while Pinecone is a managed service with plan minimums. Where each runs, what each costs, and why teams moved.
  3. ComparisonQdrant vs pgvector: A Separate Vector Database or Vectors in Postgrespgvector keeps vectors in the Postgres you run, while Qdrant runs as its own database with filters built into its index. Where each fits, what each costs, and what teams report moving between them.
Read the research