pgvector−29 Negative

Operator Insights

What are teams learning from building on pgvector and running it in production?

6 accounts: 1 failure, incident or rollback, 2 left it or turned it down and 3 switched

Failures, incidents and rollbacks1

Aashi DuttSenior technical content writer, writing Mem0's accountMem0
Mem0's agent memory service, with vectors and relational data in Aurora PostgreSQL and multi-tenant filters on every query.
Mem0 blog, published Jul 30, 2026Read the post (opens Mem0 blog in a new tab)
Observation
Mem0's HNSW indexes failed to serve its production queries: the Postgres planner chose 8-to-14-second sequential scans over a 1,153 GB table once selective tenant filters entered, and Mem0 moved its vectors to turbopuffer.
Takeaway
Check with EXPLAIN that filtered vector queries use the index at production size. Mem0 says teams under roughly 10 million vectors with simple queries probably should not migrate.
What happened
Mem0's HNSW indexes recorded 0 scans in production while every vector query fell back to an 8-to-14-second sequential scan over a 1,153 GB table. Mem0's own diagnosis: the Postgres query planner correctly decided a sequential scan was cheaper than the HNSW index once selective multi-tenant filters entered the query, and the post says this is not a pgvector bug and that the planner was doing its job. The unused indexes cost 325 GB of storage, and Mem0 moved its vectors to turbopuffer and kept relational data in Postgres, where retrieval now averages about 110 to 120 ms.
- Public organization write-up (Mem0's own blog)
- Measured results reported
- Published Jul 30, 2026, updated Sep 28, 2026
Deployment details
Mem0's agent memory service, with vectors and relational data in Aurora PostgreSQL and multi-tenant filters on every query.
- Public organization write-up (Mem0's own blog)
- Measured results reported
- Published Jul 30, 2026, updated Sep 28, 2026
Team
Mem0
Workload
Multi-tenant vector search for agent memory over a 1,153 GB table
Product version
pgvector HNSW indexes on Aurora PostgreSQL
Measurement
0 index scans; 8 to 14 second sequential scans; 325 GB of unused index storage

Left it or turned it down2

Nicolas MautiMLOps EngineerMalt
Malt's freelancer recommendation system, matching projects against over 700,000 freelancers at the time of writing (July 2024).
Malt engineering blog, published Jul 10, 2024Read the write-up (opens Malt engineering blog in a new tab)
Observation
Malt benchmarked pgvector for its freelancer matching in August 2023 and chose Qdrant instead: pgvector reached 14 queries per second at 0.40 precision against Qdrant's 90 at 0.84.
Takeaway
Benchmark on your own data and settings: Malt's figures date from August 2023, and its post notes some databases could have improved since.
What happened
Malt benchmarked candidate vector databases on the GIST1M corpus of 1,000,000 vectors of 960 dimensions in August 2023. Its post says pgvector offered an intriguing blend of relational and vector features, but its performance and ANN quality did not meet Malt's benchmarks for speed and precision: 14 queries per second at 0.40 precision, against 90 at 0.84 for Qdrant. The post notes some databases could have improved since.
- Public engineering write-up
- Measured results reported
- Published Jul 10, 2024
Deployment details
Malt's freelancer recommendation system, matching projects against over 700,000 freelancers at the time of writing (July 2024).
- Public engineering write-up
- Measured results reported
- Published Jul 10, 2024
Team
Malt engineering
Duration
Benchmark run in August 2023
Workload
Freelancer recommendation for project matching
Measurement
GIST1M benchmark, August 2023: pgvector 14 queries per second at 0.40 precision, Qdrant 90 at 0.84

GitLabSep 10, 2024GitLab, gitlab-org/gitlabGitLab dropped the embeddings database it had built on Postgres with pgvector, merging the change on Sep 10, 2024 so that it uses only its Google Vertex database.
Nathan WeinshenkerBackend Engineer, AI EngineeringGitLab
GitLab's embedding databases for its AI chat features, set up with the pgvector extension on GitLab's own Postgres infrastructure.
GitLab, gitlab-org/gitlab, published Sep 10, 2024Read the merge request (opens GitLab, gitlab-org/gitlab in a new tab)
Observation
GitLab dropped the embeddings database it had built on Postgres with pgvector, merging the change on Sep 10, 2024 so that it uses only its Google Vertex database.
Takeaway
When another managed store already covers the workload, count the upkeep of a separate pgvector database before keeping both.
What happened
GitLab's production change records show its embedding databases were set up with the pgvector extension, installed from source at first and later from a package, with errors in the AI chat bot as the stated risk of changes. A merge request opened in June 2024 and merged Sep 10, 2024 says the embeddings database is no longer in use and drops it so that GitLab solely uses its Google Vertex database. A later change request decommissioned the database.
Public engineering write-up (GitLab merge request and change records)
Outcome stated: database dropped
Merged Sep 10, 2024
Deployment details
GitLab's embedding databases for its AI chat features, set up with the pgvector extension on GitLab's own Postgres infrastructure.
Public engineering write-up (GitLab merge request and change records)
Outcome stated: database dropped
Merged Sep 10, 2024
Team
GitLab AI Engineering
Duration
Until September 2024
Workload
Embedding database for GitLab's AI chat
Product version
Postgres with the pgvector extension

Switched to or from it3

Ankit MittalEngineerInstacart
Instacart's search retrieval across retailer catalogs, rebuilt on Postgres.
Instacart tech blog, published May 28, 2025Read the write-up (opens Instacart tech blog in a new tab)
Observation
Instacart moved semantic search from FAISS to pgvector, and a production A/B test showed a 6% drop in searches with zero results thanks to better recall.
Takeaway
Prove throughput on a lab cluster that mimics production before migrating, and measure recall as well as speed, since recall is where Instacart gained.
What happened
To migrate from FAISS to pgvector, Instacart first built a lab-scale cluster that mimicked production traffic to confirm pgvector could handle its throughput and latency. Offline, pgvector was marginally slower than FAISS for larger retailers and had better recall. In a production A/B test, searches with zero results dropped 6%, and keeping retrieval in one datastore let Instacart filter documents by attributes, which its FAISS setup could not.
- Public engineering write-up
- Measured results reported
- Published May 28, 2025
Deployment details
Instacart's search retrieval across retailer catalogs, rebuilt on Postgres.
- Public engineering write-up
- Measured results reported
- Published May 28, 2025
Team
Instacart search
Workload
Semantic search retrieval with attribute filtering across retailer catalogs
Product version
pgvector in Postgres, after a lab-scale prototype cluster
Measurement
6% drop in searches with zero results in a production A/B test

Confident AIOct 29, 2023Confident AI blogConfident AI replaced Pinecone entirely with pgvector after weeks of experimentation, because Pinecone forced a complicated architecture beside its existing data.
Jeffrey IpCo-founderConfident AI
Confident AI, which builds open-source LLM evaluation infrastructure and lists a team size of 7 on its Y Combinator page, clustering LLM observability log data in production.
Confident AI blog, published Oct 29, 2023Read the post (opens Confident AI blog in a new tab)
Observation
Confident AI replaced Pinecone entirely with pgvector after weeks of experimentation, because Pinecone forced a complicated architecture beside its existing data.
Takeaway
If your vector search runs over existing, single-sourced data, Confident AI advises a data store with a built-in vector option over a standalone vector database.
What happened
Confident AI first used Pinecone to cluster LLM observability log data in production. After weeks of experimentation it replaced Pinecone entirely with pgvector, writing that Pinecone's design hides complexities in integrating with existing data storage and that its restrictive metadata storage made it troublesome.
Public organization write-up (co-founder's post on the company blog)
Outcome stated
Published Oct 29, 2023
Deployment details
Confident AI, which builds open-source LLM evaluation infrastructure and lists a team size of 7 on its Y Combinator page, clustering LLM observability log data in production.
Public organization write-up (co-founder's post on the company blog)
Outcome stated
Published Oct 29, 2023
Team
Confident AI
Workload
Clustering LLM observability log data in production
Center for AI Innovation, NCSAJul 27, 2026GitHub, Center-for-AI-Innovation/Illinois-ChatNCSA's Center for AI Innovation replaced Qdrant with pgvector in Illinois Chat, expecting lower cost and easier upkeep in the Postgres instance it already ran.
Luigi MariniLead Research Software Engineer, National Center for Supercomputing Applications (NCSA)Center for AI Innovation, NCSA
Illinois Chat, the Center for AI Innovation's chat application at the University of Illinois Urbana-Champaign, which ran Qdrant on one VM beside a Postgres instance.
GitHub, Center-for-AI-Innovation/Illinois-Chat, published Jul 27, 2026Read the issue (opens GitHub, Center-for-AI-Innovation/Illinois-Chat in a new tab)
Observation
NCSA's Center for AI Innovation replaced Qdrant with pgvector in Illinois Chat, expecting lower cost and easier upkeep in the Postgres instance it already ran.
Takeaway
Moving to pgvector may need a Postgres image or host that ships the extension, as NCSA found with the official Postgres image.
What happened
The issue Replace qdrant with pgvector, opened Jul 27, 2026 as a public copy of an internal issue, says Qdrant is difficult to maintain and incredibly costly on one VM and asks whether the current Postgres instance supports pgvector. A pull request removed Qdrant and switched to the official pgvector image, since the official Postgres image does not include the extension, and was merged Aug 10, 2026.
Public organization write-up (issue in the center's own GitHub repository)
Outcome stated: issue closed as completed
Opened Jul 27, 2026
Deployment details
Illinois Chat, the Center for AI Innovation's chat application at the University of Illinois Urbana-Champaign, which ran Qdrant on one VM beside a Postgres instance.
Public organization write-up (issue in the center's own GitHub repository)
Outcome stated: issue closed as completed
Opened Jul 27, 2026
Team
Center for AI Innovation, NCSA, University of Illinois Urbana-Champaign
Workload
Illinois Chat's vector store
Product version
The official pgvector image, replacing Qdrant on one VM

Failures, incidents and rollbacks1

Mem0's HNSW indexes failed to serve its production queries: the Postgres planner chose 8-to-14-second sequential scans over a 1,153 GB table once selective tenant filters entered, and Mem0 moved its vectors to turbopuffer.

Left it or turned it down2

Malt benchmarked pgvector for its freelancer matching in August 2023 and chose Qdrant instead: pgvector reached 14 queries per second at 0.40 precision against Qdrant's 90 at 0.84.

GitLab dropped the embeddings database it had built on Postgres with pgvector, merging the change on Sep 10, 2024 so that it uses only its Google Vertex database.

Switched to or from it3

Instacart moved semantic search from FAISS to pgvector, and a production A/B test showed a 6% drop in searches with zero results thanks to better recall.

Confident AI replaced Pinecone entirely with pgvector after weeks of experimentation, because Pinecone forced a complicated architecture beside its existing data.

NCSA's Center for AI Innovation replaced Qdrant with pgvector in Illinois Chat, expecting lower cost and easier upkeep in the Postgres instance it already ran.