Softwr

Machine Learning · head to head

Seldon vs turbopuffer

Seldon logo

Seldon

Machine Learning

Kubernetes model serving whose current version is licensed under the Business Source Licence

From
Free
Free plan
Yes
turbopuffer logo

turbopuffer

Databases

Closed-source vector and full-text search service built directly on object storage, with cold queries measured in seconds rather than milliseconds.

From
$16/month
Free plan
No

The short version

  • Seldon is free; turbopuffer starts at $16/mo.
  • Only Seldon lists kubernetes custom resources; only turbopuffer lists object storage architecture.
  • Prices and features on this page were last checked on 13 November 2026.

Where they differ

Only the attributes on which Seldon and turbopuffer actually diverge.

Attributes where Seldon and turbopuffer differ
AttributeSeldonturbopuffer
Starting priceFree$16/month
Pricing modelFreemiumSubscription
Free tierYesNo
PlatformsLinuxWeb
CategoryMachine LearningDatabases

Not yet recorded: founded for turbopuffer.

What each one covers

Drawn from each product's published feature list. An absence here means we hold no record of it - not that the product lacks it.

Only in Seldon

  • Kubernetes custom resources
  • Inference graphs
  • Traffic strategies
  • Open Inference Protocol
  • Alibi Explain
  • Alibi Detect
  • Kafka-backed pipelines in v2
  • Commercial control plane

Only in turbopuffer

  • Object storage architecture
  • Namespaces
  • Vector search
  • Full-text search
  • Attribute filtering
  • Documented limits
  • Configurable consistency
  • Durable writes

What people use each for

The jobs each tool is most often brought in to do.

Seldon

  • Serving an ensemble or a multi-stage inference path as one versioned deployment rather than as a chain of separate servicesnot turbopuffer
  • Running genuine production experiments where a share of live traffic goes to a candidate model and the results are comparednot turbopuffer
  • Regulated environments needing explanations and drift monitoring attached to the served model rather than bolted on laternot turbopuffer
  • Organisations with an established Kubernetes platform team who want serving expressed as manifests under existing deployment controlsnot turbopuffer

turbopuffer

  • A product with one search index per customer and thousands of customers, most of whose data is idle on any given daynot Seldon
  • Very large corpora where holding every vector in memory is the dominant cost and occasional cold-query latency is acceptablenot Seldon
  • Hybrid retrieval combining BM25 and vector search where running and synchronising two separate systems is the problem being solvednot Seldon
  • Retrieval for agent and assistant products where indexes are created and destroyed frequently and per-index overhead must be near zeronot Seldon

Where each one falls short

Documented limitations, not opinions. Every one is a constraint you would hit in normal use.

Seldon

  • Seldon Core v2 is under the Business Source Licence rather than Apache 2.0, so production use requires a commercial agreement, and a team that evaluated it believing it was open source discovers the licence is the blocker exactly when the project is ready to ship
  • Core v1 remains Apache 2.0 but is in maintenance, so taking the free route means running software that receives no new development while the architecture it belongs to moves on without it
  • Version 2 is a different system rather than a newer release, with different custom resources, a scheduler component and a Kafka-based pipeline model, so migrating from v1 is a re-implementation of every deployment manifest rather than an upgrade
  • Kafka is a dependency for v2 pipelines, so an organisation that does not already operate it takes on a distributed log with its own storage, retention, rebalancing and failure modes purely in order to serve models
  • Everything assumes Kubernetes fluency and the failure modes are Kubernetes failure modes, custom resource version mismatches, an operator that will not reconcile, admission webhooks and resource limits terminating an inference pod mid-request, so it needs a platform engineer rather than a data scientist

turbopuffer

  • A cold namespace pays object storage latency on the first query, with a documented p90 around 1,214 ms on a million documents, so any interactive search box needs the data kept warm or the user waits about a second
  • Queries are eventually consistent by default, and after roughly 128 MiB of outstanding writes new data is invisible until indexed, which the vendor puts at tens of seconds for small namespaces and tens of minutes for large ones, so a bulk re-index is not immediately queryable
  • It is closed source with no community edition, so single-tenant or bring-your-own-cloud deployment is a commercial negotiation rather than a deployment choice, and there is no path to running it yourself if the relationship ends
  • Per-namespace ceilings, roughly 10,000 writes per second, 32 MB/s and 500 million documents per shard, mean a single enormous index has to be sharded across namespaces by your application rather than by the service
  • It is a search engine, not a database: there are no joins, no cross-document transactions and no SQL, so it sits beside a primary datastore and keeping the two in step is work that belongs to you

Pricing, plan by plan

Seldon

Free
  • Seldon CoreFree
    • Open source
    • Kubernetes deployment
    • Model serving
  • Seldon DeployFree
    • Enterprise features
    • GUI
    • Monitoring

turbopuffer

$16/month
  • Launch$16/mo
    • All database features
    • Multi-tenancy deployment
    • SOC2 & GDPR-ready DPA
  • Scale$256/mo
    • Everything in Launch
    • HIPAA-ready BAA
    • Single Sign-On (SSO)
  • Enterprise$4,096/mo
    • Everything in Scale
    • Single-tenancy & BYOC deployment options
    • Private networking

Which should you pick?

Choose Seldon if

  • You need kubernetes custom resources.
  • You want to start without paying.
  • You work on Linux.
  • You also want inference graphs.

Choose turbopuffer if

  • You need object storage architecture.
  • You also want namespaces.

Questions people ask

Is Seldon or turbopuffer better?
Neither clearly leads. Seldon starts at Free and turbopuffer at $16/month, and user ratings are close enough to be indistinguishable. Choose on capability and platform support.
Which is cheaper, Seldon or turbopuffer?
Seldon has a free tier; the other does not. Paid plans start at Free for Seldon and $16/month for turbopuffer.
Does Seldon or turbopuffer run on more platforms?
Seldon runs on Linux. turbopuffer runs on Web.
Can I use Seldon for free?
Yes. Seldon has a free tier, so you can try it without paying. turbopuffer starts at $16/month.
What is Seldon best used for?
Seldon is most often used for serving an ensemble or a multi-stage inference path as one versioned deployment rather than as a chain of separate services, running genuine production experiments where a share of live traffic goes to a candidate model and the results are compared, regulated environments needing explanations and drift monitoring attached to the served model rather than bolted on later, organisations with an established kubernetes platform team who want serving expressed as manifests under existing deployment controls. Of those, serving an ensemble or a multi-stage inference path as one versioned deployment rather than as a chain of separate services and running genuine production experiments where a share of live traffic goes to a candidate model and the results are compared are not what turbopuffer is typically brought in for.
What can Seldon do that turbopuffer cannot?
Seldon covers Kubernetes custom resources, Inference graphs, Traffic strategies, Open Inference Protocol. turbopuffer covers Object storage architecture, Namespaces, Vector search, Full-text search.

Answered from the vendors’ own pages

Seldon: Is Seldon open source?

Partly, and this is the thing to check before you build on it. Core v1 is Apache 2.0 but in maintenance. Core v2 was moved to the Business Source Licence in 2024, which allows evaluation but not unlicensed production use. Verify the current licence of each component you intend to run, including MLServer and the Alibi libraries.

turbopuffer: Can I self-host turbopuffer?

There is no open source or community edition. Single-tenant and bring-your-own-cloud deployments exist as commercial arrangements, but there is no way to run it independently of the vendor.

Seldon: What is the difference between v1 and v2?

Architecture, not just version number. v2 introduces a scheduler, a different set of custom resources and Kafka-backed pipelines. Manifests, mental model and operations all change, so treat a move as a project.

turbopuffer: How fast is it really?

Warm queries perform comparably to in-memory search engines. Cold queries, where data is not cached, have a documented p90 around 1,214 ms on a million documents. Write p90 is around 248 ms for a 512 KB upsert because writes go straight to object storage.

Seldon: Do I need Kubernetes?

Yes. It is a Kubernetes-native system and there is no meaningful deployment without a cluster and someone competent to run it.

turbopuffer: Is it consistent?

Eventually consistent by default, with the vendor reporting that over 99.8% of queries return consistent data. Strong consistency can be requested per query at a latency cost. Large write bursts have a longer visibility delay while indexing catches up.

Seldon: What is MLServer?

Seldon's Python inference server implementing the Open Inference Protocol, usable inside Seldon deployments or on its own. Check its current licence alongside Core's, since the company has moved projects onto the Business Source Licence.

turbopuffer: What is it best at?

Large numbers of namespaces where most are idle. The architecture makes cold data cheap to keep, which is exactly the shape of a multi-tenant product with a long tail of inactive customers.

Seldon: Do I have to run Kafka?

For v2 pipelines, yes. If you only need single models served, that dependency is a large amount of infrastructure for the benefit, and a simpler serving layer may be the better answer.

turbopuffer: What are the hard limits?

Up to 128 billion documents and 256 TB per namespace, 500 million documents per shard, 64 MiB per document, 10,752 dense vector dimensions, roughly 10,000 writes per second per namespace and a maximum result set of 10,000.

Share

Related pages

Other head to heads