Qdrant−40 Negative

Operator Insights

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

4 accounts: 2 failures, incidents or rollbacks, 1 left it or turned it down and 1 adopted it

Failures, incidents and rollbacks2

Javier MartinezSoftware EngineerZylon
Zylon, the company behind PrivateGPT, whose settings default the vector store to Qdrant, running Qdrant on Kubernetes with the official Helm chart and image.
GitHub, qdrant/qdrant, published Oct 16, 2024Read the issue (opens GitHub, qdrant/qdrant in a new tab)
Observation
Zylon's production Qdrant pods failed under ingestion as memory grew until Kubernetes evicted them. After moving payload indexes to disk they held at 512Mi and still crashed at 256Mi.
Takeaway
Set every part of the collection you want off RAM to on-disk, payload indexes included, and load-test at the memory limit you plan to run.
What happened
In October 2024 Zylon's engineer reported that memory grew with each ingestion batch until the pod reached its limit, leading to frequent evictions and restarts in what the engineer called their production configuration, which already set vectors and HNSW on disk. Qdrant replied that memory use follows the configuration, and that on-disk data is released under memory pressure. After putting payload indexes on disk as well, the pods ran fine at a 512Mi limit with 40 concurrent users and still crashed at 256Mi.
- Public engineering write-up (issue by a Zylon engineer)
- Measured results reported
- Opened Oct 16, 2024
Deployment details
Zylon, the company behind PrivateGPT, whose settings default the vector store to Qdrant, running Qdrant on Kubernetes with the official Helm chart and image.
- Public engineering write-up (issue by a Zylon engineer)
- Measured results reported
- Opened Oct 16, 2024
Team
Zylon
Workload
Large-scale data ingestion into Qdrant with llama-index
Product version
Tested with 0.10.0 and 1.12.1 (official Helm chart and Docker image) on Kubernetes 1.30.0
Measurement
Held at a 512Mi memory limit with 40 concurrent users and crashed at 256Mi

MaltDec 11, 2023GitHub, qdrant/qdrantMalt's cluster upgrade from Qdrant 1.6.0 to 1.7.1 failed to run without downtime because new pods would not start beside old ones, so Malt held the production upgrade for Qdrant's fix.
Nicolas MautiMLOps EngineerMalt
Malt's self-hosted Qdrant cluster on Kubernetes, deployed with the official Helm chart.
GitHub, qdrant/qdrant, published Dec 11, 2023Read the issue (opens GitHub, qdrant/qdrant in a new tab)
Observation
Malt's cluster upgrade from Qdrant 1.6.0 to 1.7.1 failed to run without downtime because new pods would not start beside old ones, so Malt held the production upgrade for Qdrant's fix.
Takeaway
Rehearse each Qdrant upgrade on a staging cluster with the same Helm chart, and hold production until the release you need is out.
What happened
On Dec 11, 2023 Malt's engineer reported that when upgrading a Qdrant cluster from 1.6.0 to 1.7.1, the new pods could not start correctly while the old ones were still running, so the upgrade could not be done without downtime. A Qdrant maintainer confirmed the issue the same day, promised a patch and suggested disabling the readiness probe during the upgrade. Malt's engineer replied that the integration cluster would get that workaround while production waited for the fix.
Public engineering write-up (issue by a Malt engineer)
Outcome stated: production upgrade held
Opened Dec 11, 2023
Deployment details
Malt's self-hosted Qdrant cluster on Kubernetes, deployed with the official Helm chart.
Public engineering write-up (issue by a Malt engineer)
Outcome stated: production upgrade held
Opened Dec 11, 2023
Team
Malt
Workload
Rolling cluster upgrade on Kubernetes
Product version
Qdrant 1.6.0 to 1.7.1 with the official Helm chart

Left it or turned it down1

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 dropped Qdrant as Illinois Chat's default store for pgvector, calling Qdrant difficult to maintain and incredibly costly on one VM.
Takeaway
For a single-VM deployment that already runs Postgres, weigh the upkeep of a second database against what Qdrant adds before you adopt it.
What happened
In the issue Replace qdrant with pgvector, opened Jul 27, 2026 as a public copy of an internal issue, Luigi Marini wrote that Qdrant is difficult to maintain and incredibly costly on one VM, and that pgvector should bring cost down and make it easier to maintain. The change landed through a pull request that removed Qdrant and used the pgvector image, merged Aug 10, 2026, and the issue was closed as completed on Aug 17, 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
Qdrant on one VM, replaced by pgvector

Adopted it1

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 chose Qdrant for its freelancer matching retriever after its own benchmark and self-hosts it on Kubernetes, and p95 matching latency fell from tens of seconds to 3 seconds at most.
Takeaway
Run your own benchmark on your filters before choosing, as Malt did, and note that its benchmark dates from August 2023.
What happened
Malt benchmarked vector databases in August 2023 and chose Qdrant for speed, accuracy and filtering, especially geo-spatial filters. It self-hosts Qdrant on its Kubernetes clusters with Qdrant's Docker image and Helm chart, and in 2023 deployed a retriever-ranker architecture with the Qdrant-backed neural retriever ahead of its matching model. The post reports p95 latency dropping from tens of seconds, sometimes over a minute, to 3 seconds at most, with recommendation quality confirmed by higher project conversion in an A/B test.
- 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
Deployed in 2023
Workload
Freelancer recommendation for project matching, over 700,000 freelancers
Product version
Self-hosted on Kubernetes with Qdrant's Docker image and Helm chart
Measurement
p95 matching latency from tens of seconds to 3 seconds at most

Failures, incidents and rollbacks2

Zylon's production Qdrant pods failed under ingestion as memory grew until Kubernetes evicted them. After moving payload indexes to disk they held at 512Mi and still crashed at 256Mi.

Malt's cluster upgrade from Qdrant 1.6.0 to 1.7.1 failed to run without downtime because new pods would not start beside old ones, so Malt held the production upgrade for Qdrant's fix.

Left it or turned it down1

NCSA's Center for AI Innovation dropped Qdrant as Illinois Chat's default store for pgvector, calling Qdrant difficult to maintain and incredibly costly on one VM.

Adopted it1

Malt chose Qdrant for its freelancer matching retriever after its own benchmark and self-hosts it on Kubernetes, and p95 matching latency fell from tens of seconds to 3 seconds at most.