Key findings
- Filtered search turns fast once fields are indexed: Manfred Dreese reports P99s falling from hundreds of milliseconds to single digits (April 2026).
- Memory and disk draw many reports: deletes need free disk to reclaim space, a full disk crashed nodes around 1 in 6 attempts in one cluster (October 2026), and recovery hit Kubernetes memory limits.
- Upgrades regressed in public reports, among them 1.16.1 and 1.18.3, which later releases fixed.
- One team moving to pgvector named keeping Postgres and Qdrant in sync as the reason (March 2025), and NCSA's Center for AI Innovation moved Illinois Chat to pgvector in August 2026, citing upkeep and cost on one VM.
Public accounts of Qdrant in production praise its filtered search and low upkeep, and spend most of their words on memory and disk sizing, upgrades and cluster recovery. This piece reads 48 public accounts from 2025 and 2026 and three named deployments against Qdrant's own documentation, stored on November 1, 2026.
Where Qdrant earns its place
Malt chose Qdrant for its freelancer matching after benchmarking vector databases in August 2023, and reports p95 latency falling from tens of seconds to 3 seconds at most. A tech lead running Qdrant for three years reports P99s in single digits once filters were indexed, and quantization cutting memory by over 4x with negligible recall impact (April 2026).
Smaller teams praise low upkeep: one runs many collections of millions of points in production, another picked Qdrant over Milvus for simplicity, and a third pays 250 EUR a month for all its Qdrant instances across all environments of a 300,000-product catalog (January 2026).
Where teams spend their time
Memory and disk come first. Deleting 5 million points freed only about 10 GiB on one 33-million-point deployment whose low disk had stopped optimization (December 2025), and recovery that hits the Kubernetes memory limit loops in CrashLoopBackOff until limits are raised by hand (April 2026).
Upgrades come next. A refactor in 1.18.2 moved HNSW links to heap memory, and on 1.18.3 memory roughly doubled for one team until 1.19.0 (October 2026), while 1.16.1 made one user's test fail until 1.16.2 a week later (November 2025).
When teams leave
The common exit is consolidation into Postgres. An entrepreneur was moving everything to pgvector because keeping Postgres and Qdrant in sync was a pain (March 2025), and NCSA's Center for AI Innovation replaced Qdrant with pgvector in Illinois Chat in August 2026, citing upkeep and cost on one VM. One developer left for LanceDB, an embedded store, when the app needed no running server (July 2026).
Methodology
GILD built this piece from 48 practitioner accounts dated 2025 and 2026 (GitHub issues and a pull request, Hacker News comments, Reddit comments and a post, PeerSpot reviews and a practitioner blog) and three named deployment accounts, each checked by a separate checker against its stored page, and from Qdrant's own pages stored on November 1, 2026. Figures are the authors' own, and none is a GILD benchmark.
Products in this piece
Sources
- Official documentation
- 4
- Public source
- 7
- Qdrant on GitHub: README
- Qdrant: Pricing
- Qdrant docs: Overview
- Qdrant docs: Resource quotas
- Malt engineering blog: a vector database for freelancer recommendation (Jul 10, 2024)
- Manfred Dreese: three years with Qdrant (Apr 23, 2026)
- qdrant/qdrant issue #9986: memory and latency after 1.18.3 (Jul 27, 2026)
- qdrant/qdrant issue #7813: disk use after deleting points (Dec 21, 2025)
- Center-for-AI-Innovation/Illinois-Chat issue #97: replace Qdrant with pgvector (Jul 27, 2026)
- Hacker News, bobosha: moving from Qdrant to pgvector (Mar 18, 2025)
- Hacker News, p410n3: 250 EUR a month for 300,000 products (Jan 9, 2026)