Skip to main content
QdrantProduct page

Qdrant Signals

What teams building AI features on Qdrant report, with every finding tied to its evidence.

Read the findings

What the evidence says

Whose sources count

Practitioner sentiment, teams and leaders

−40 Negative from 41 sources that take a side

How this figure is made

Negative overall signal, net 40 points negative

From 40 authors in 37 discussions. Run a pilot on your own documents and filters to measure recall against exact search, query latency at your filter selectivity, and memory per million vectors.

Analyzed
72 total sources
Top praise
Low Operational Work (6 of 7 sources) See praise evidence
Top concern
Cluster Operations (9 of 9 sources) See concern evidence

Jan 2025 to Oct 2026.

48 sources from practitioners, by month and sentiment.

Teams running Qdrant clusters report fast filtered retrieval once payload indexes are in place, and recurring work on memory sizing, upgrades and cluster recovery. Regressions in 1.16.1 and 1.18.3 needed a later release or a downgrade.

Positive, mixed and negative sources

The score counts them as 7 positive, 10 mixed and 23 negative votes, one per author. See How the numbers work.

See the evidence

Below: 1 praise theme, 4 concerns and 1 disagreement.

Findings

Counted for Teams and technical leaders, from 41 sources. Change group

What practitioners praise

Every capability practitioners describe positively, ranked by how many sources say so. Open one to read who said it.

  1. Low Operational Work6 of 7 sources on it praise itTeams report that Qdrant, self-hosted or managed, takes little ongoing operational work.

    Counted for Teams and technical leaders.

    Positive, 6

    1. Positive

      sinevilson, Reddit commenter

      Reddit comment by sinevilsonReddit comment

      Runs Qdrant in production with many collections over millions of points, and says the operations take work once, maybe twice, to automate everything.

      Published Jul 15, 2026self-described: Running Qdrant in production with many collections

    2. Positive

      Samay Kapadia (sa-code), Principal ML Engineer, Zalando (GitHub profile linked from HN)

      Hacker News comment by sa-codeHacker News comment

      Calls Qdrant a good default choice because the same engine runs in memory for development, on a hard drive for small deployments, and for large workloads.

      Published Jan 25, 2026self-described: Setting a default vector store for a team

    3. Positive

      couAUIA, Hacker News commenter

      Hacker News comment by couAUIAHacker News comment

      Lists Qdrant under what worked for the company's internal search: easy to use, solid performance, and well suited to self-hosting.

      Published Nov 5, 2025self-described: An engineering company building internal document search

    4. Positive

      eujzmc, Reddit commenter

      Reddit comment by eujzmcReddit comment

      The team picked Qdrant for simplicity and would only switch to Milvus if scale really demanded it.

      Published Sep 24, 2025self-described: A SaaS team with millions of vectors choosing between Milvus and Qdrant

    5. Positive

      junegirl_198, Reddit commenter

      Reddit comment by junegirl_198Reddit comment

      Self-hosted Qdrant is lightweight, easy to manage and much cheaper than Pinecone or Zilliz as you scale.

      Published Sep 3, 2025self-described: A production deployment with lots of ingests and lower query volume

    6. Positive

      k9294, Hacker News commenter

      Hacker News comment by k9294: A relatively small startup running managed QdrantHacker News comment

      Managed Qdrant for vector search works well, is quite simple and scales well. It needed an API boundary layer and more work upfront, and was quite low-maintenance in the long run.

      Published Mar 17, 2025self-described: A relatively small startup running managed Qdrant

    Mixed or negative, 1

    1. Negative

      k9294, Hacker News commenter

      Hacker News comment by k9294: Comparing Qdrant and Zilliz Cloud against turbopuffer on operational simplicityHacker News comment

      Tried Qdrant and Zilliz Cloud and says turbopuffer is ahead of both on operational simplicity.

      Published Sep 9, 2025self-described: Comparing Qdrant and Zilliz Cloud against turbopuffer on operational simplicity

    GILD's fact checks, which don't count as sources

    Checked Nov 1, 2026.Qdrant runs self-hosted on 64-bit x86_64 or ARM64 systems and needs block-level storage with a POSIX-compatible file system. It does not work with NFS or S3. For testing or development it runs as a container or binary, and the Kubernetes Helm chart has community support only. Self-hosted installation

    Checked Nov 1, 2026.Qdrant Managed Cloud runs natively on AWS, GCP and Azure. 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. Managed Cloud

    Checked Nov 1, 2026.Qdrant Cloud registration works with an email address and passwordless login, a Google account, a GitHub account, or an enterprise SSO connection. A free tier cluster needs no credit card. Cloud account setup

    Open them in the record

What concerns them

Every concern practitioners raise, most raised first.

  1. Cluster OperationsAll 9 sources on it raise it as a concernRecovery, snapshots, networking, storage and node authentication in multi-node or Kubernetes deployments need care.

    Counted for Teams and technical leaders.

    1. Negative

      R.Sicart (rsicart), JobTeaser (GitHub profile)

      GitHub issue by rsicartGitHub issue

      Random reads and writes stalled for about 62 s before returning 200, which the reporter traced to idle internal gRPC connections dropped by the network. The issue was closed by a linked change.

      Published Oct 20, 2026self-described: A 3-node Qdrant Hybrid Cloud cluster on Kubernetes with Calico eBPF

    2. Negative

      Benedikt Volkmer (bvolkmer), Machine Learning Researcher and Developer (GitHub profile)

      GitHub issue by bvolkmerGitHub issue

      The Python client's get_snapshot always fails, because snapshots are tar files and the wrapper parses JSON. A maintainer replied that the clients do not expose it and to download snapshots through the raw REST API.

      Published Aug 19, 2026self-described: Downloading snapshots through the Python client

    3. Mixed

      Manfred Dreese, tech lead and consultant (his own blog)

      Practitioner blog post by Manfred DreesePractitioner blog post

      Once his team indexed its filters properly, P99s dropped from hundreds of milliseconds to single digits, and quantization cut the memory footprint by over 4x with negligible impact on recall. When WAL recovery itself hits the Kubernetes memory limit, the pod is killed again and ends in a CrashLoopBackOff until someone bumps the limits by hand.

      Published Apr 23, 2026self-described: Three years running Qdrant for a large-scale recommendation product on Kubernetes

    4. Negative

      Frank Wall (fraenki), markt.de (GitHub profile)

      GitHub issue by fraenkiGitHub issue

      Raised that nothing stopped unauthorized nodes from joining a cluster. Qdrant's self-hosted security page now says that as of v1.20.0 the receiving peer checks the key by default, and that without an API key no authentication layer is installed.

      Published Oct 27, 2025self-described: Operating a self-hosted Qdrant cluster

    5. Negative

      Rauf Akdemir (orhanrauf), co-founder and CTO, Airweave (GitHub profile)

      GitHub issue by orhanraufGitHub issue

      After a pod restart, shard recovery ran at 3 items/second while using 1 CPU core of 12 and 1.1GB of 100GB RAM. The issue was open when stored.

      Published Aug 23, 2025self-described: Self-hosting Qdrant v1.15.3 on Kubernetes (Azure AKS) with a 50,000-document collection

    6. Negative

      Konrad Mosoń (morsik), DevOps, openhades (GitHub profile)

      GitHub issue by morsikGitHub issue

      Nodes returned data inconsistently. A maintainer said points created in v1.15.0 might be related to a known bug, which the team confirmed. 4 of 6 nodes had indexed the collection on that version, so fixing it meant replacing about 1/3 of the data.

      Published Aug 18, 2025self-described: A 6-node Qdrant cluster

    7. Negative

      Aalekh Patel (aalekhpatel07), Canadian Centre for Child Protection (GitHub profile)

      GitHub issue by aalekhpatel07GitHub issue

      Recovering a collection from its snapshot loses the collection's aliases. Qdrant declined to change the snapshot and advises migrating aliases with a separate request.

      Published Jun 24, 2025self-described: Restoring collections with aliases from snapshots

    8. Negative

      Rajan Patil (rajanpatil), Elsevier (GitHub profile)

      GitHub issue by rajanpatilGitHub issue

      Snapshot recovery failed with Azure File Share storage. A Qdrant maintainer replied that Azure File Share is not supported and that data corruption is possible later even where collections seem to work.

      Published Mar 28, 2025self-described: A single Qdrant node on Azure Kubernetes Service

    9. Negative

      Robert Deckel (bdeckel), PTC (GitHub profile)

      GitHub issue by bdeckelGitHub issue

      An upgrade from 1.13.4 to 1.13.5 failed and the pipeline rolled back, after the team's own deployment had added a step that removed raft_state.json. A maintainer said the file should never be removed, and advised putting the original back if kept, otherwise rebuilding the cluster.

      Published Mar 21, 2025self-described: A single Qdrant pod on Kubernetes in Microsoft Azure with the default Helm values

    GILD's fact checks, which don't count as sources

    Checked Nov 1, 2026.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. Qdrant Cloud SLA

    Open them in the record
  2. Memory and Disk SizingAll 7 sources on it raise it as a concernMemory and disk needs run higher than planned, and running short stalls or crashes nodes.

    Counted for Teams and technical leaders.

    1. Negative

      dstockton (David Stockton), GitHub issue reporter

      GitHub issue by dstocktonGitHub issue

      After raising replication beyond what the disks could hold, around 1 in 6 attempts crashed a node, with or without the disk quota enabled. A fix was still open when the page was stored.

      Published Oct 23, 2026self-described: A 3-node cluster after raising a collection's replication factor

    2. Mixed

      Manfred Dreese, tech lead and consultant (his own blog)

      Practitioner blog post by Manfred DreesePractitioner blog post

      Once his team indexed its filters properly, P99s dropped from hundreds of milliseconds to single digits, and quantization cut the memory footprint by over 4x with negligible impact on recall. When WAL recovery itself hits the Kubernetes memory limit, the pod is killed again and ends in a CrashLoopBackOff until someone bumps the limits by hand.

      Published Apr 23, 2026self-described: Three years running Qdrant for a large-scale recommendation product on Kubernetes

    3. Negative

      Deniz Özdemirli (02deno), Computer Engineer at Nokia (GitHub profile)

      GitHub issue by 02denoGitHub issue

      Queries that worked before the upgrade made Qdrant 1.16.3 restart with a panic and then OOMKilled errors. A Qdrant maintainer, timvisee, read the jump in memory as more than 55GB.

      Published Jan 23, 2026self-described: Hybrid RRF queries with min_should filters on Qdrant 1.16.3

    4. Negative

      ffletcherr (Homayoun), GitHub issue reporter

      GitHub issue by ffletcherrGitHub issue

      With 325 GiB used for about 33 million points, removing 5 million points freed only about 10 GiB, and deleting 5 million more pushed disk usage to 100%. A maintainer replied that Qdrant needs headroom on disk to work.

      Published Dec 21, 2025self-described: A Kubernetes deployment with tens of millions of points

    5. Negative

      ml-abdula (Abdullah), GitHub issue reporter

      GitHub issue by ml-abdulaGitHub issue

      Queries took around 7 seconds with everything on disk, against a target under 1 second. A maintainer said Qdrant normally recommends keeping quantized vectors and HNSW in RAM.

      Published Sep 3, 2025self-described: A production deployment of about 4.6M points with vectors and HNSW on disk

    6. Negative

      amircodota, GitHub issue reporter

      GitHub issue by amircodotaGitHub issue

      After the team moved vectors to disk, disk usage in the production database grew fast until it ran out of space, with 492G in temp_segments against 67G of segments. A maintainer advised upgrading from 1.8.4 through each minor version.

      Published Jun 3, 2025self-described: A large production collection on Qdrant 1.8.4

    7. Negative

      Mohammad Saknini (Mohammadsaknini), Data & AI Engineer at Inform GmbH (GitHub profile)

      GitHub issue by MohammadsakniniGitHub issue

      During a bulk upload toward 89 million points in Docker on WSL, the container started at roughly 7GB of RAM and kept climbing to the reporter's 30GB limit after 5 to 6 million points, with vectors and HNSW on disk. The reporter found it happened only under Docker, and a maintainer recommended a Linux native machine for production.

      Published Jan 8, 2025self-described: Bulk ingestion of 89 million 768-dimension points in Docker on WSL, for research

    GILD's fact checks, which don't count as sources

    Checked Nov 1, 2026.Qdrant has built-in quantization that cuts RAM usage by up to 97%, with a tunable trade-off between search speed and precision. Quantization

    Checked Nov 1, 2026.A free tier cluster has 1 node with 1 GB RAM, 0.5 vCPU and 4 GB disk, which supports about 1 M vectors of 768 dimensions. It runs without high availability, with only manual snapshots and restores via API and version upgrades with downtime. Unused free clusters are suspended after 1 week and deleted after 4 weeks of inactivity if not reactivated. Free Tier limits

    Checked Nov 1, 2026.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. Vector limits

    Checked Nov 1, 2026.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. Resource quotas

    Open them in the record
  3. Upgrade RegressionsAll 6 sources on it raise it as a concernMinor and patch upgrades bring memory, data or startup problems that need a staged path or a later fix.

    Counted for Teams and technical leaders.

    1. Negative

      Jakob Edding (jkbe), bakdata (GitHub profile)

      GitHub issue by jkbeGitHub issue

      The HNSW links residency bug led to about 2x memory consumption under v1.18.3 compared to v1.17.1, and his team persistently ran into OOM on startup, which made v1.18.3 a broken release for them.

      Published Oct 13, 2026self-described: A memory-scarce Qdrant deployment with only the HNSW index in memory

    2. Negative

      ronenperez, GitHub issue reporter

      GitHub issue by ronenperezGitHub issue

      After upgrading 1.18.1 to 1.18.3, each node gained 19 to 25 GB of heap memory and latency regressed, and downgrading to v1.18.1 fully reversed both. A maintainer closed the issue as implemented in Qdrant 1.19.0.

      Published Jul 27, 2026self-described: A production Qdrant cluster with on-disk vectors

    3. Mixed

      Kieran (bytestream), SupportPal (GitHub profile)

      GitHub issue by bytestreamGitHub issue

      The staged upgrade from 1.10.1 went smoothly until the final step, where the 1.16.3 Linux release binary could not open storage that still relied on RocksDB. The maintainer's workaround was to start 1.16.3 once from the official Docker image.

      Published Feb 19, 2026self-described: A staged upgrade from Qdrant 1.10.1 to 1.16.3

    4. Negative

      Deniz Özdemirli (02deno), Computer Engineer at Nokia (GitHub profile)

      GitHub issue by 02denoGitHub issue

      Queries that worked before the upgrade made Qdrant 1.16.3 restart with a panic and then OOMKilled errors. A Qdrant maintainer, timvisee, read the jump in memory as more than 55GB.

      Published Jan 23, 2026self-described: Hybrid RRF queries with min_should filters on Qdrant 1.16.3

    5. Mixed

      Boris Petrov (boris-petrov), profuz (GitHub profile)

      GitHub issue by boris-petrovGitHub issue

      Upgrading from 1.16.0 to 1.16.1 made one of the reporter's tests fail consistently. Qdrant released 1.16.2 with the fix a week later, and the reporter thanked the team for the quick fix.

      Published Nov 27, 2025self-described: Upgrading self-hosted Qdrant with an automated test suite

    6. Negative

      Konrad Mosoń (morsik), DevOps, openhades (GitHub profile)

      GitHub issue by morsikGitHub issue

      Nodes returned data inconsistently. A maintainer said points created in v1.15.0 might be related to a known bug, which the team confirmed. 4 of 6 nodes had indexed the collection on that version, so fixing it meant replacing about 1/3 of the data.

      Published Aug 18, 2025self-described: A 6-node Qdrant cluster

    Open them in the record
  4. A Separate Server to RunAll 5 sources on it raise it as a concernRunning Qdrant as its own server adds sync work, setup or latency that an embedded store or the existing database avoids.

    Counted for Teams and technical leaders.

    1. Negative

      SilverConsistent9222, Reddit poster

      Reddit post by SilverConsistent9222Reddit post

      Wasted time setting up Qdrant way too early, when a local Chroma was completely fine for the corpus the team had.

      Published Jun 7, 2026self-described: An internal knowledge base tool built at work

    2. Negative

      Amrit Dash, PeerSpot reviewer

      PeerSpot review by Amrit DashPeerSpot review

      Running Supabase and Qdrant meant every operation updated both, which risked the two falling out of sync when an API call failed and complicated the developers' work.

      Published May 26, 2026self-described: Automation engineer at an educational organization with 11 to 50 employees, per PeerSpot's reviewer line

    3. Mixed

      Vedant Wani, PeerSpot reviewer

      PeerSpot review by Vedant WaniPeerSpot review

      Recommendation accuracy rose from 45 to 50% with FAISS to around 85 to 95% with Qdrant. The local file lock forced the team to run Qdrant as an always-on Docker service.

      Published May 18, 2026self-described: Analyst at Synergy Connect, per PeerSpot's reviewer line

    4. Negative

      Arlo Gilbert (arlogilbert), serial founder (HN profile), Osano (GitHub)

      Hacker News comment by arlogilbertHacker News comment

      On 3 text collections of 10k items each, averaged over about 500 queries per collection, Qdrant answered in 21.1ms against 5.9ms for LanceDB and 0.8ms for Zvec. The author ran Qdrant and LanceDB with Inverse Document Frequency enabled and calls that a slight performance hit.

      Published Feb 23, 2026self-described: Comparing Zvec, LanceDB and Qdrant on a small text dataset

    5. Mixed

      bobosha, 3x entrepreneur and AI researcher (HN profile)

      Hacker News comment by boboshaHacker News comment

      Ran Qdrant in production and recommends it, but the team is moving everything to Postgres with pgvector for fewer moving parts, because keeping data synced between Postgres and Qdrant was a pain.

      Published Mar 18, 2025self-described: A team that ran Qdrant in production beside Postgres

    Open them in the record

Where the evidence agrees

Statements this group's sources agree on, then the ones close to agreement.

For teams and technical leaders, no statement reaches 5 sources from 3 authors.

Where it disagrees

Questions the sources split on, with the count on each side.

  • Evidence remains split

    Is running Qdrant light operational work?

    • Light, 5 sources

      Teams running Qdrant in production with many collections, or picking it over Milvus, report little ongoing work.

    • Heavy, 3 sources

      Others report memory limits, crash loops, slow shard recovery or full disks that took manual work.

The evidence

Whose experience this is

The three groups, then every practitioner by the context they work in, and how emphasis changes with scale.

Three groups drawn from the same sources. Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. Signals open on Teams and technical leaders.

  • Teams and technical leaders

    41 sources. Named teams and founders describe Qdrant in production on Kubernetes and Qdrant Cloud.

  • Individual builders

    7 sources. Individual developers describe Qdrant for local search, RAG and agent memory.

  • All practitioners

    48 sources. Together, these accounts combine individual developers with platform teams running clusters.

48 sources from practitioners, by the context they work in. Filled rows count for this group.

  1. Individual builderDevelopers building on the platform on their own7 sources

    Emphasizes Local search and RAG on one machine, and whether a server is needed at all

  2. Small product team2 to 10 engineers, or an unnamed startup team7 sources, counted here

    Emphasizes Low maintenance and price against Pinecone, and the cost of syncing a second store

  3. Product or platform team11 to 50 engineers, or a named company's team26 sources, counted here

    Emphasizes Memory and disk sizing, upgrade regressions and cluster operations on Kubernetes

  4. Technical leadershipLeaders setting model and platform choices across teams6 sources, counted here

    Emphasizes Founders and tech leads on filter indexing, recall at scale and recovery time

  5. Organization-wideEnterprise or company-wide AI platform2 sources, counted here

    Emphasizes Company approval of FastEmbed: its license shown wrongly on PyPI and a package firewall's license finding. The Operator Insights carry Malt, Zylon and NCSA.

What changes with scale. Individual developers weigh whether a separate server is worth running. Teams on Kubernetes report memory, disk and recovery work, and founders and tech leads report fast filtered search once fields are indexed.

Where the sources come from

Where this group's sources were published.

Counted for Teams and technical leaders, from 41 sources. Change group

65 sources for this group, 24 of them documentation

  • Developer Communities36, 55% of 65 sources
  • Product Documentation24, 37% of 65 sources
  • Review Sites4, 6% of 65 sources
  • Practitioner Blogs1, 2% of 65 sources

Earlier evidence remains when it still applies. A bug report on a release that Qdrant later fixed is marked Still applicable, since the upgrade lesson holds.

See sources by origin

The record

Who each source is, how it knows, whether it still applies and what it counts toward.

Newest first, on its own page. Every count above opens its sources there.

Read all 72 sources

About this read

What changed, and how we count

What changed

  • Nov 1, 2026: first edition, built from 48 practitioner records and 36 documentation statements that a different model checked against their stored pages.
  • Every account links to its public source and names its author and publication date.
  • A finding shows only when at least 5 sources from 3 different authors discuss it. A disagreement also requires at least 2 sources on each side.

Method

A running review of what teams building AI features on Qdrant report, from individual builders to company-wide platforms. Each finding links to its sources, and each conflict shows both sides.

Counts cover only the sources GILD reviewed.

Who counts

Teams and technical leaders
Small product team, Product or platform team, Technical leadership, Organization-wide
Individual builders
Individual builder
All practitioners
Everyone in the teams and leaders and individual builders groups

Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. GILD files each source in a group from how the account describes its author. A role or employer marked self-described is what the source states about themselves. Documentation checks facts for every group and never counts toward sentiment.

How the numbers work

  • Overall signal. Positive votes minus negative ones, per 100 votes from the sources that take a side. Each author counts once, with their latest view, whoever they are and wherever they posted. The older weighted reading, where accounts GILD matched to a named person count three times when GILD lists the author among technical leaders and twice otherwise, is shown beside it as a second view. Where an author wrote more than one source, the line beside the split gives the votes, one per author, so the score follows from it. When one author's sources from the same day disagree, that author counts as mixed, and a half point rounds away from zero. Sources with no opinion are listed and not counted. Strong positive is +50 or higher, positive +20 to +49, mixed −19 to +19, negative −20 to −49, strong negative −50 or lower.
  • Weighted reading, second view. Counted the older way, where an account GILD matched to a named person counts two or three times, teams and leaders read −37 negative (method sentiment/S1). GILD printed that number as "GILD score" until October 29, 2026; most of those matches rest on a name or a self-described role.
  • Unscored. A group gets a score only with 20 sources from 15 authors. Below that, one source moves the score about 5 points, so the read says "Unscored until 20 sources from 15 authors" and shows its counts.
  • Independence. A source whose author works for the vendor, for one of its partners or programs, or for a competitor, and a page the vendor published or ran, is listed with that tie and counts toward nothing. One author counts on at most two records per product. The line under the score says how many authors and how many separate discussions the counted sources come from: comments in one thread are one discussion.
  • Praise and concerns. How many of the sources that discuss a theme praise it or raise it.
  • Consensus. A statement joins when 20% of the group's sources support it and fewer contradict it. Contradicting sources are counted separately. Below that line a statement shows as building. At 40% or more it reads Strong agreement; below, Recurring point.
  • Disagreements. The count on each side. GILD names no winner and gives no share of all users.
  • Minimum evidence. A finding needs 5 sources from 3 authors; a disagreement also needs 2 on each side. Smaller counts stay in the table below, marked under the minimum.

Labels on a source

A source's date and whether it still applies are separate: a recent post can describe an old version, and an old one can still hold.

Documented
Stated in official documentation.
Reported
A practitioner's account of their own use.
Measured
Backed by the source's own measurements, such as a comparative test or tracked team metrics.
GILD analysis
GILD's check of a claim against official documentation.
Current
Describes the product as it is now.
Still applicable
Older, and the behavior or lesson still holds.
Needs a recheck
May depend on a plan, version or setting that changed since it was written. It keeps counting until GILD rechecks it, since the rest of the account usually holds, and its card says so.
Out of date
No longer accurate. Kept in the record; counts toward nothing current.

Flagged for a recheck, of the sources that take a side: teams and leaders 0 of 41, individual builders 0 of 7, all practitioners 0 of 48.

Limits

  • These figures describe reviewed sources and do not represent all Qdrant users. GitHub issues and pull requests supply 26 of the 48 practitioner sources, Hacker News comments 9, Reddit comments and posts 8, PeerSpot reviews 4 and one practitioner's blog 1.
  • Sources run from January 2025 to October 2026. GitHub reports name the Qdrant release they ran, from 1.8.4 to 1.20.0, and most Hacker News, Reddit and PeerSpot accounts name none.

Every count, and the sources behind it

14 findings for 3 groups

Each row opens its sources for the group you are reading.

FindingTeams and technical leadersIndividual buildersAll practitionersSources
Current readCurrent sentiment8 positive, 10 mixed, 23 negative2 positive, 2 mixed, 3 negative10 positive, 12 mixed, 26 negativeSources
PraiseLow Operational Work6 of 7 positive1 of 1 positive (under the minimum)7 of 8 positiveSources
PraiseCost Against Hosted Services3 of 3 positive (under the minimum)0 of 0 positive (under the minimum)3 of 3 positive (under the minimum)Sources
PraiseFramework Integrations0 of 0 positive (under the minimum)1 of 1 positive (under the minimum)1 of 1 positive (under the minimum)Sources
PraiseEmbeddings and Model Options0 of 2 positive (under the minimum)0 of 0 positive (under the minimum)0 of 2 positive (under the minimum)Sources
PraiseSecurity and Compliance0 of 2 positive (under the minimum)0 of 0 positive (under the minimum)0 of 2 positive (under the minimum)Sources
ConcernMemory and Disk Sizing7 of 7 raise it2 of 2 raise it (under the minimum)9 of 9 raise itSources
ConcernUpgrade Regressions6 of 6 raise it0 of 0 raise it (under the minimum)6 of 6 raise itSources
ConcernCluster Operations9 of 9 raise it1 of 1 raise it (under the minimum)10 of 10 raise itSources
ConcernRecall and Throughput Under Filters4 of 4 raise it (under the minimum)0 of 0 raise it (under the minimum)4 of 4 raise it (under the minimum)Sources
ConcernA Separate Server to Run5 of 5 raise it1 of 2 raise it (under the minimum)6 of 7 raise itSources
DisagreementIs running Qdrant light operational work?Light 5, Heavy 3Light 0, Heavy 1 (under the minimum)Light 5, Heavy 4Sources
Current readEvidence mix65 sources31 sources72 sourcesSources

Fact checks

GILD's checks of claims against official documentation, kept apart from the source counts.

  • Low Operational Work. Qdrant runs self-hosted on 64-bit x86_64 or ARM64 systems and needs block-level storage with a POSIX-compatible file system. It does not work with NFS or S3. For testing or development it runs as a container or binary, and the Kubernetes Helm chart has community support only. Self-hosted installation, checked Nov 1, 2026
  • Low Operational Work. Qdrant Managed Cloud runs natively on AWS, GCP and Azure. 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. Managed Cloud, checked Nov 1, 2026
  • Low Operational Work. Qdrant Cloud registration works with an email address and passwordless login, a Google account, a GitHub account, or an enterprise SSO connection. A free tier cluster needs no credit card. Cloud account setup, checked Nov 1, 2026
  • Search Speed and Accuracy. Qdrant supports dense vectors for semantic similarity, sparse vectors for full-text search, and multivector search for objects with multiple embeddings or late interaction models such as ColBERT. Vector types, checked Nov 1, 2026
  • Search Speed and Accuracy. Qdrant attaches a JSON payload to vectors and filters on it with keyword matching, full-text, numeric range and geo-location conditions, combined with should, must and must_not clauses. Payload filtering, checked Nov 1, 2026
  • Search Speed and Accuracy. Qdrant combines multiple vectors in a single query and merges the results with fusion strategies such as Reciprocal Rank Fusion (RRF) and Distribution-Based Score Fusion (DBSF). Hybrid search, checked Nov 1, 2026
  • Search Speed and Accuracy. Qdrant scales horizontally with sharding and replication and updates or resizes collections with zero downtime. Distributed deployment, checked Nov 1, 2026
  • Cost Against Hosted Services. 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. Pricing: Free Tier, checked Nov 1, 2026
  • Cost Against Hosted Services. Qdrant Cloud Standard has usage-based pricing. Billing follows actual resource usage of compute (vCPU), memory (GB), cluster storage (GB), backup storage (GB) and used inference tokens of paid models, and usage is billed hourly. Pricing: Standard metering, checked Nov 1, 2026
  • Cost Against Hosted Services. 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. Marketplace billing, checked Nov 1, 2026
  • Cost Against Hosted Services. 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. Premium plan terms, checked Nov 1, 2026
  • Cost Against Hosted Services. Qdrant is licensed under the Apache License, Version 2.0, which grants a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license. License, checked Nov 1, 2026
  • Framework Integrations. LangChain distributes the Qdrant integration as the langchain-qdrant partner package, which supports dense, sparse and hybrid retrieval. LangChain integration, checked Nov 1, 2026
  • Framework Integrations. Qdrant's framework integrations list includes LlamaIndex, a data framework for building LLM applications. Framework integrations, checked Nov 1, 2026
  • Framework Integrations. The Haystack integration is QdrantDocumentStore, a separate qdrant-haystack package that the Qdrant team maintains and updates independently of Haystack. Haystack integration, checked Nov 1, 2026
  • Framework Integrations. Qdrant provides a REST API with an OpenAPI 3.0 specification and a gRPC interface. On Qdrant Cloud, REST runs on port 6333 or 443 and gRPC on 6334, and clients exist for Python, TypeScript, Rust, Go, Java and .NET. API and clients, checked Nov 1, 2026
  • Embeddings and Model Options. Qdrant stores and searches vectors from any embedding model. A dense vector can have between 1 and 65,536 dimensions. Embeddings: bring your own vectors, checked Nov 1, 2026
  • Embeddings and Model Options. Managed deployments on Qdrant Cloud have access to Cloud Inference, where Qdrant Cloud hosts a range of embedding models, some for free. BM25 and all-MiniLM-L6-v2 are free to use, including on free-tier clusters. Cloud Inference is a Managed Cloud feature, and Hybrid Cloud and self-hosted deployments use client-side inference or in-cluster BM25. Cloud Inference: hosted models, checked Nov 1, 2026
  • Embeddings and Model Options. On Qdrant Cloud, Cloud Inference can call embedding models hosted by OpenAI, Cohere, Jina AI and OpenRouter, which requires your own API key for the provider. Cloud Inference: external providers, checked Nov 1, 2026
  • Embeddings and Model Options. FastEmbed is Qdrant's lightweight Python library for embedding generation. Self-hosted Qdrant uses client-side inference, for example with FastEmbed, and the Python client runs inference locally with FastEmbed when cloud inference is off. FastEmbed, checked Nov 1, 2026
  • Embeddings and Model Options. Cloud Inference is billed per token at a fixed, region-independent price that depends on the model, shown in the Cloud Console. Free models carry no token charges. On paid clusters, paid models run on metered tokens after a monthly free allowance of up to 5 million tokens per model. Cloud Inference billing, checked Nov 1, 2026
  • Embeddings and Model Options. Each Cloud Inference 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. Cloud Inference input limits, checked Nov 1, 2026
  • Security and Compliance. On Qdrant Cloud, data in Qdrant clusters is stored only in the cluster's deployment region, and telemetry and logs are pushed to Qdrant's US management plane. Data residency on Qdrant Cloud, checked Nov 1, 2026
  • Security and Compliance. 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. Self-hosted audit logging, checked Nov 1, 2026
  • Security and Compliance. 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. SOC 2 Type 2, checked Nov 1, 2026
  • Security and Compliance. Qdrant is HIPAA certified, and a Business Associate Agreement is available for Qdrant Managed Cloud. PHI handling requirements are discussed with Qdrant. HIPAA, checked Nov 1, 2026
  • Security and Compliance. For customers subject to GDPR, Qdrant offers a Data Processing Agreement that covers its data protection and privacy commitments. GDPR data processing agreement, checked Nov 1, 2026
  • Security and Compliance. 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, which it destroys when that obligation expires. Service agreement: data deletion, checked Nov 1, 2026
  • Security and Compliance. Qdrant Cloud backup schedules set Days of Retention, the number of days after which a backup snapshot is deleted. Cloud backups: retention, checked Nov 1, 2026
  • Security and Compliance. Under the Qdrant Cloud Service Agreement, Qdrant uses Customer data on Managed Cloud to make it available to Team Members and to provide the Cloud Service and support services. Qdrant accesses Customer data only to support the customer's use and address service or technical problems, to render other Qdrant Services, or as the customer permits in writing. The agreement lets Qdrant use Aggregated Data, which does not identify the customer, for developing and improving its products. Service agreement: data use, checked Nov 1, 2026
  • Security and Compliance. 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. Qdrant Cloud data access and encryption, checked Nov 1, 2026
  • Memory and Disk Sizing. Qdrant has built-in quantization that cuts RAM usage by up to 97%, with a tunable trade-off between search speed and precision. Quantization, checked Nov 1, 2026
  • Memory and Disk Sizing. A free tier cluster has 1 node with 1 GB RAM, 0.5 vCPU and 4 GB disk, which supports about 1 M vectors of 768 dimensions. It runs without high availability, with only manual snapshots and restores via API and version upgrades with downtime. Unused free clusters are suspended after 1 week and deleted after 4 weeks of inactivity if not reactivated. Free Tier limits, checked Nov 1, 2026
  • Memory and Disk Sizing. 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. Vector limits, checked Nov 1, 2026
  • Memory and Disk Sizing. 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. Resource quotas, checked Nov 1, 2026
  • Cluster Operations. 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. Qdrant Cloud SLA, checked Nov 1, 2026