Skip to main content

Decision brief

Qdrant Review: A Dedicated Vector Database, Self-Hosted or on Qdrant Cloud

Qdrant filters vectors on JSON payloads with hybrid search and quantization, under Apache 2.0 or on Qdrant Cloud. Plan payload indexes, memory and disk headroom, and staged upgrades before production.

Reading time
5 min read
Sources
33 cited
In this piece
Qdrant−40 Negative

GILD verdict

Choose Qdrant when retrieval needs its own database with selective metadata filters, and your team can run it on Kubernetes or pay for Qdrant Cloud.

Create payload indexes for every filtered field before ingesting, plan memory and disk with headroom, and stage upgrades through each minor release. If your data already lives in Postgres and filters are simple, pgvector saves you a second store to keep in sync.

Qdrant

Vector Database by Qdrant Solutions GmbH

−40 Negative
Product page

Fit and limits

Who it fits

  • Queries combine vector search with selective filters: Qdrant filters on JSON payloads with keyword, full-text, numeric range and geo conditions.
  • You want one engine from development to production, self-hosted under Apache 2.0 or on Qdrant Cloud on AWS, GCP and Azure.
  • Memory is the constraint: built-in quantization cuts RAM usage by up to 97%, per Qdrant's README.
  • You want hosted embeddings beside the database: Qdrant Cloud Inference hosts embedding models, and BM25 and all-MiniLM-L6-v2 are free on free-tier clusters.

Where it falls short

  • Teams that keep application data in Postgres or Supabase report that keeping a second store in sync was a pain or a risk, and one was moving everything to pgvector (March 2025 and May 2026).
  • Memory and disk needs run high: deletes do not free disk until the optimizer runs, and a full disk crashed nodes in one cluster around 1 in 6 attempts (October 2026).
  • Patch releases have regressed: 1.18.3 roughly doubled memory for one team until 1.19.0 (October 2026), and 1.16.1 made one user's test fail until 1.16.2 (November 2025).
  • Single sign-on needs the Premium tier and covers Cloud Console sign-in only, and GILD found no SCIM provisioning on Qdrant's pages.

Implementation reality

  1. Model the payload first

    Decide which fields you filter on and create payload indexes for them before ingesting, so the filterable HNSW index can add filter-aware edges. Nested filters on arrays of objects need an extra step that reads the original payload, so flatten fields you filter on often.

  2. Size memory and disk

    By default Qdrant stores vectors in RAM. Move original vectors to disk and keep quantized vectors and HNSW in RAM, and leave disk headroom for deletes, optimizer runs and recovery. From v1.19.0, resource quotas can stop a node from taking more data when memory or disk runs low.

  3. Pick where it runs

    Self-hosted Qdrant needs block-level storage with a POSIX file system and does not work with NFS or S3. Qdrant Cloud runs on AWS, GCP and Azure with data stored in the cluster's region. Hybrid Cloud runs in your own infrastructure and the Private Cloud Enterprise Operator runs on Kubernetes without a cloud connection, both by contacting Qdrant.

  4. Stage upgrades

    Upgrade through each minor release with a test run on staging. Public reports show regressions in 1.16.1 and 1.18.3 that later releases fixed, and a release binary that needed one start from the official Docker image.

What to verify before you buy

Plans and billing

  • Free Tier: The Qdrant Cloud Free Tier is free forever, for testing and prototypes, with a single node cluster of 0.5 vCPU, 1 GB RAM and 4 GB disk.
  • Standard and Premium: Standard is usage-based, paid as you go or under a committed-spend contract. Premium uses Premium tier rates under a committed-spend contract with an annual minimum.
  • Marketplace: Through an AWS, GCP or Azure Marketplace subscription, Qdrant Cloud usage is added as US$0.01 Resource Usage Units, so a cluster costing US$85 in a month adds 8,500 units to the cloud provider bill. Pricing does not differ between payment methods.
  • Inference: Free models carry no token charges.

Identity, security and data

  • Certifications: Qdrant holds SOC 2 Type 2, with compliance reports in its Trust Center for Qdrant Cloud and for Hybrid and Private Cloud. The security page marks compliance documentation as not applicable to self-hosted Qdrant. Qdrant is HIPAA certified, and a Business Associate Agreement is available for Qdrant Managed Cloud.
  • GDPR: For customers subject to GDPR, Qdrant offers a Data Processing Agreement that covers its data protection and privacy commitments.
  • Audit logs, self-hosted: In self-hosted Qdrant, audit logging is available from v1.17.0, records API operations that require authentication or authorization as JSON, and is not enabled by default. By default, audit logs rotate daily and the seven most recent log files are kept.
  • Encryption and access: On Qdrant Cloud, every storage volume is encrypted at rest, Qdrant does not access data stored in Qdrant clusters, and API keys are stored as hashes. Premium customers can encrypt data at rest with their own keys.
  • Retention: Customers can delete any Customer data from the Managed Cloud Service at any time during the term. Qdrant deletes or destroys Customer Data 30 days after the Cloud Service Order expires, unless a statutory obligation requires it to retain some. Backup schedules set Days of Retention, the number of days after which a backup snapshot is deleted.
  • Data use: The service agreement lets Qdrant use Aggregated Data, which does not identify the customer, for developing and improving its products.
  • SCIM and training: GILD read 89 Qdrant deploy, inference, security, pricing and legal pages on Nov 1, 2026 and found no statement of SCIM provisioning, or of whether Qdrant trains models on customer data.

Limits and uptime

  • Free clusters: A free tier cluster supports about 1 M vectors of 768 dimensions and runs without high availability, with only manual snapshots and restores via API and version upgrades with downtime.
  • Vectors: A dense vector can have between 1 and 65,536 dimensions. For multivectors, the number of sub-vectors times the vector size must stay under 1,048,576, a limit that is not configurable.
  • Cloud Inference input: Each model has a context window, and longer input text is truncated to fit. For dense models, the collection's vector size must match the model output, or the upsert fails with an error.
  • SLA: The Qdrant Cloud SLA commits to quarterly uptime of 99.5% for Standard, 99.9% for Premium single-AZ clusters and 99.95% for Premium multi-AZ clusters. It covers Managed Cloud clusters only, and Hybrid Cloud and Private Cloud fall outside it.
  • Quotas: From Qdrant v1.19.0, resource quotas stop a node from taking on more data once it is low on memory or disk space, and quotas are disabled by default.

Deployment

  • Self-hosted: For testing or development it runs as a container or binary, and the Kubernetes Helm chart has community support only.
  • Managed Cloud regions: Free clusters have a limited choice of cloud providers and regions. Paying through a cloud marketplace limits cluster creation to paid clusters in that provider's regions.

What to measure in your pilot

  • No stored Qdrant page prints a per-unit price for Standard compute, memory or storage, or a token rate for paid inference models. Price your cluster in the Cloud calculator or Console before committing.
  • Qdrant's own pages differ on whether single sign-on is a Premium add-on or included in Premium, so confirm it in your order.
  • The one public account of recall under a tenant filter (August 2025, recurring in February 2026) predates later releases, so measure filtered recall against exact search on your own data.

Sources

Official documentation
30
Public source
3
  1. Qdrant: PricingOfficial documentationqdrant.tech
  2. Qdrant on GitHub: READMEOfficial documentationraw.githubusercontent.com
  3. Qdrant on GitHub: LICENSEOfficial documentationraw.githubusercontent.com
  4. Qdrant docs: Create a clusterOfficial documentationqdrant.tech
  5. Qdrant docs: Pricing and paymentsOfficial documentationqdrant.tech
  6. Qdrant docs: Premium tierOfficial documentationqdrant.tech
  7. Qdrant docs: Hybrid CloudOfficial documentationqdrant.tech
  8. Qdrant docs: Private CloudOfficial documentationqdrant.tech
  9. Qdrant docs: EmbeddingsOfficial documentationqdrant.tech
  10. Qdrant docs: VectorsOfficial documentationqdrant.tech
  11. Qdrant docs: InferenceOfficial documentationqdrant.tech
  12. Qdrant docs: Cloud InferenceOfficial documentationqdrant.tech
  13. Qdrant: Cloud InferenceOfficial documentationqdrant.tech
  14. Qdrant docs: FastEmbedOfficial documentationqdrant.tech
  15. Qdrant docs: Cloud account setupOfficial documentationqdrant.tech
  16. Qdrant: llms.txtOfficial documentationqdrant.tech
  17. Qdrant docs: LangChainOfficial documentationqdrant.tech
  18. Qdrant docs: FrameworksOfficial documentationqdrant.tech
  19. Qdrant docs: HaystackOfficial documentationqdrant.tech
  20. Qdrant docs: InstallationOfficial documentationqdrant.tech
  21. Qdrant docs: Managed CloudOfficial documentationqdrant.tech
  22. Qdrant: SecurityOfficial documentationqdrant.tech
  23. Qdrant docs: Configure a clusterOfficial documentationqdrant.tech
  24. Qdrant docs: Security (self-hosted)Official documentationqdrant.tech
  25. Qdrant blog: SOC 2 Type II auditOfficial documentationqdrant.tech
  26. Qdrant docs: Cloud securityOfficial documentationqdrant.tech
  27. Qdrant Cloud Service AgreementOfficial documentationcloud.qdrant.io
  28. Qdrant docs: Cloud backupsOfficial documentationqdrant.tech
  29. Qdrant Cloud SLAOfficial documentationcloud.qdrant.io
  30. Qdrant docs: Resource quotasOfficial documentationqdrant.tech
  31. Hacker News, bobosha: moving from Qdrant to pgvector (Mar 18, 2025)Public sourcenews.ycombinator.com
  32. qdrant/qdrant issue #9986: memory and latency after 1.18.3 (Jul 27, 2026)Public sourcegithub.com
  33. qdrant/qdrant issue #7813: disk use after deleting points (Dec 21, 2025)Public sourcegithub.com

Keep reading

  1. Buyer guideQdrant Production PlaybookA checklist for taking Qdrant to production: payload indexes before ingest, memory and disk headroom, supported storage, and staged upgrades.
  2. ResearchQdrant in Production: What Teams Report in 2026Where Qdrant earns its place in public accounts, and where teams spend their time: memory and disk sizing, upgrades and cluster recovery.
  3. ComparisonPinecone vs Qdrant: Managed Only or Self-HostablePinecone runs only as a managed service, while Qdrant also runs on your own servers and ARM machines. Where each runs, how each prices its entry plans, and what teams report.
Read the research