Qdrant Trial
Overview
Qdrant is an open source vector database with filtering, hybrid search, and horizontal scaling options for RAG and recommendation workloads (Qdrant docs). It is written in Rust, licensed under Apache 2.0, and available self-hosted, as the managed Qdrant Cloud, or as Hybrid Cloud running in the customer's own Kubernetes cluster (Qdrant GitHub, Qdrant Cloud).
Trial as a dedicated vector tier when pgvector or warehouse-native search lacks latency or filtering requirements. Plan collection governance, backup, and embedding version strategy alongside adoption.
Adoption Signals
- Funding and usage: Qdrant announced a USD 50 million Series B led by AVP on 12 March 2026, with Bosch Ventures participating, and reported more than 250 million downloads and named users including Canva, HubSpot, Roche, Bosch and Tripadvisor (Qdrant Series B).
- Regulated-sector and German enterprise use: Deutsche Telekom built LMOS, the multi-agent platform behind its Frag Magenta assistants, on Qdrant and reports cutting new-agent development from 15 days to 2 (Deutsche Telekom case study).
- Sovereign deployment pattern: FAZ runs Qdrant Hybrid Cloud on Azure Kubernetes Service with more than 14 million vectors and sub-second full-archive searches, keeping infrastructure in its own environment (FAZ case study).
- Active development: v1.19.0 (5 August 2026) added 4-bit TurboQuant storage, per-component memory modes and a global quota API; v1.19.1 followed on 4 September 2026; the repository had about 34,900 GitHub stars on 30 September 2026 (v1.19.0 release notes, Qdrant GitHub).
- Ecosystem reach: Feast supports Qdrant as an online store, so feature and retrieval pipelines can share it (Feast GitHub).
Risks
- Insecure defaults when self-hosted. The security guide states that self-deployed instances are open to all network interfaces with no authentication and are "not production-ready" until API keys, TLS and binding are configured (Qdrant security).
- Stepwise upgrades. Self-hosted clusters must pass through the latest patch of every intermediate minor version, and single-node or replication-factor-1 collections need downtime during upgrades (Qdrant upgrades).
- API deprecations. v1.19.0 removed deprecated search endpoints from the OpenAPI specification and deprecated a strict-mode memory setting (v1.19.0 release notes); client code written against older examples needs review.
- Embedding lifecycle. A new embedding model means re-embedding and re-indexing whole collections; without versioned collections and aliases, model changes turn into outages or silent relevance regressions.
- Vendor-reported outcomes. The Telekom and FAZ figures come from Qdrant's own case studies and should be validated with your own recall and latency tests.
Pros & Cons
Advantages
- European production references with control over data location: Deutsche Telekom uses Qdrant as the vector layer of its LMOS multi-agent platform, and FAZ searches its 75-year archive with Qdrant Hybrid Cloud on its own Azure Kubernetes cluster.
- Strong filtered and hybrid retrieval in one engine: payload filters, dense and sparse vectors, multi-vector representations and quantization options such as the 4-bit TurboQuant storage added in v1.19.0.
- Apache-2.0 core with a funded company behind it: Qdrant raised a USD 50 million Series B in March 2026 and reports more than 250 million downloads, while self-hosting remains possible without a commercial licence.
Disadvantages
- Self-hosted Qdrant is not secure by default: it listens on all interfaces without authentication, so API keys, JWT-based access control, TLS and network binding must be configured before any production use.
- Self-hosted upgrades must go one minor version at a time, and cluster nodes are only compatible across one minor version, which turns skipped upgrades into multi-step migrations.
- It adds a second stateful system next to the system of record, with its own sharding, replication, snapshots and embedding re-indexing, which pgvector or warehouse-native search avoids for smaller workloads.
Recommendation
Trial Qdrant when pgvector or warehouse-native vector search no longer meets requirements: heavily filtered queries over tens of millions of vectors, hybrid dense-plus-sparse retrieval, or multi-tenant RAG with strict latency targets. Measure recall and p95 latency against pgvector on your own data first; for small corpora that already live in PostgreSQL, the extra system is rarely worth it.
For regulated workloads, prefer Qdrant Hybrid Cloud or a self-hosted cluster in an EU region, and treat security configuration as a go-live gate: API keys or JWT access per collection, TLS, private networking, snapshots to object storage and a tested restore. Version collections by embedding model, switch traffic with aliases, keep a replication factor of at least 2 for rolling upgrades, and plan one minor-version upgrade per release cycle so the cluster never falls more than one version behind.
Sources
Consulting
From radar to implementation
You have assessed the technologies — we help you bring them into production systems safely and measurably.
