A model trained on last month's image collection has suddenly become hard to reproduce. The training code has a Git commit. The bucket path looks familiar. Yet several labels behind that path have since been corrected. lakeFS gives the dataset a version that a training run can record, so an engineer can distinguish a change to the examples from a change to the code that learned from them.
- 01What lakeFS manages It keeps the references and metadata that describe a version of objects in a repository. The objects reside in underlying storage.
- 02How a branch starts A new branch points to an existing commit. Teams can propose and isolate dataset changes without first duplicating the entire collection.
- 03Where a check lives Actions can associate an event such as a proposed merge with validation logic. The team supplies the rule that matters for its data.
01 / ProductVersion history above the object store
A lakeFS repository groups related objects. They may be pictures, annotations, video segments or table files. A commit identifies a fixed view of that collection, while a branch provides a movable reference for ongoing work. The distinction is familiar to developers who use Git, but the objects in this case can be too large or numerous to fit sensibly into the source-code repository. lakeFS manages the references while the underlying object store holds the data itself. lakeFS platform lakeFS data model and internals
A zero-copy branch gives several teams the same starting point. When one team corrects a label file, the correction can be evaluated on its own branch. The unchanged images continue to share their original data. A new version of a large changed object can still consume storage. The advantage is that starting the experiment does not require a full physical clone of everything the team might use. lakeFS data model and internals
The company also acquired the DVC open-source project from Iterative. DVC and lakeFS have related origins in reproducible data science, but a purchase of lakeFS Enterprise is a separate decision from using an open-source project. An individual who already tracks a dataset in a Git-oriented DVC workflow has no reason to change it solely because of the acquisition. The relevant question is where different teams need to coordinate changes to shared object-store data. DVC project acquisition
02 / AudienceWhen two teams need to disagree about the same dataset
Consider an inspection team that labels photographs of manufactured parts. A training engineer needs a stable collection for the next model. A quality supervisor has discovered that one shift used the wrong label for a visible defect. Both have legitimate claims on the data. One needs to reproduce the earlier model, while the other needs to correct the record. Branches let the correction progress while the earlier run retains a precise point of reference.
A platform team has a different interest. Each group may already keep its own dated copies, but nobody knows which copy contains an approved exception or which evaluation set went with the training set. A repository can group the relevant objects and commit them together. The platform owner still has to decide what belongs inside that boundary. A training job that also reads an unversioned configuration file outside the repository has another input that should be recorded.
A collection of a few small text fixtures that change with application code may already fit in Git. At the other extreme, an organisation using cloud storage for historical records needs more than a version identifier to satisfy retention obligations. Ask how often multiple collaborators actually branch, review, merge or reproduce shared data. Those actions are a stronger justification for lakeFS than the size of a bucket by itself.
03 / WorkflowA proposed label correction with a reproducible training run
Pick one part family with an agreed set of image identifiers, annotations and holdout examples. Identify the exact starting commit. In the experiment record, write down that identifier alongside the training code revision, model settings and method for splitting examples. Two runs referencing a moving main branch can start from different data even when their configuration files appear identical.
Create a branch for the label correction. Change the relevant annotations and retain a small manifest explaining which examples the supervisor intended to change. Ask a second person to inspect a handful of the changed image-label pairs against the written definition of the defect. A perfectly valid file can still assign the wrong class. A useful human review catches that semantic error before it becomes a well-versioned mistake.
Add a narrowly defined check to a pre-merge action. For example, an annotation change might require every referenced image ID to exist and every class label to occur in an agreed vocabulary. The hooks guide describes actions files that select events and branches and send work to validation logic such as a webhook. A failed validation can stop the merge. These particular label rules are proposed checks the inspection team would implement, not built-in knowledge of its factory. Actions and validation hooks
Train the experimental model from the candidate commit and the comparison model from the earlier commit. Use the same untouched evaluation set and training procedure for both. If the score changes, pull individual examples that moved from a correct to an incorrect classification and vice versa. Reviewers can now trace a changed prediction back to the dataset revision and decide whether a labeling improvement exposed a weakness in the old evaluation set.
After approval, merge the intended change and record the accepted commit for the released model. Try reconstructing the input list from that fixed reference in a fresh environment. Include file contents for a sample of changed labels, not just names. This checks the handoff to the training system and helps establish which other inputs, such as preprocessing code or a manually downloaded vocabulary, still sit outside the versioned dataset.
04 / PricingWhat the commercial quote depends on
The community project is available in its public source repository. A team operating it needs an appropriate object store, metadata storage, deployment and maintenance. lakeFS Enterprise is available as a managed cloud service or self-managed in the cloud or on premises. The pricing page invites teams to request a free 15-day Enterprise trial. Its quote depends on requirements including users, data under management, API performance and deployment model. lakeFS architecture lakeFS Enterprise pricing lakeFS community source repository
| Route | Commercial basis | What the team must size |
|---|---|---|
| Community project | Open-source software, self-operated | Object storage, metadata service and operating effort |
| Enterprise self-managed | A quote for the deployment and requirements | Users, governed data, API workload and local operation |
| Managed enterprise service | A quote for managed delivery | Service responsibilities, data location and required performance |
Deployment and pricing routes, consulted 21 November 2026. lakeFS Enterprise pricing
For an informative quote, describe the active writers as well as the consumers. A scheduled ingest that updates many small objects stresses a different part of a system from a long-running experiment that reads a few large files repeatedly. Document the size of replaced objects, the number of candidate branches and how long old revisions must remain recoverable. A zero-copy branch can reduce duplication at creation while retained changes still have a cost.
05 / DistinctionsWhere dataset versioning meets experiment tracking
The Comet Blueprint covers another part of the training record. An experiment tracker helps people relate parameters, metrics and outputs to a run. A lakeFS commit gives that record a specific dataset reference. Linking the two is valuable when somebody asks which input labels produced a surprising result. A commit alone cannot explain optimizer settings, while a run name such as “corrected labels” cannot reconstruct the exact files.
The Databricks Blueprint is relevant if most of the organisation's work already takes place in a managed analytics environment. Determine which assets its existing processes can recover together. A table of image identifiers might be versioned while the image objects or accompanying JSON annotations remain mutable. Trace one training example across each storage boundary to see whether lakeFS would close a genuine gap.
06 / QuestionsWhich paths could bypass a reproducible dataset
List every way a training job and an annotator reach the collection. If one application reads a branch reference and another reads a mutable bucket path directly, the models can see different inputs. The architecture documentation separates physical object storage from the metadata store and serving components. Test the ordinary SDK or processing framework used by the team and record the exact reference it resolves. lakeFS architecture
Decide who can approve a validation rule change and who handles a blocked release when the checking service is unavailable. A rule that rejects unknown labels is useful until the product legitimately acquires another defect class. Rehearse introducing the new class, changing the rule and recovering the previous approved dataset. That exercise reveals whether data owners can maintain their own contract or need a platform engineer for every exception.
07 / DecisionThree decisions for the team that owns the data
Reproducing a disputed model result
Take one existing training record and attempt to recover its examples. Record a fixed dataset commit in the next run and repeat the recovery with a colleague.
Reviewing corrected training labels
Create an isolated branch, make the smallest coherent correction and pair an automated reference check with human inspection of changed examples.
Choosing a deployment
Describe who owns the object store, metadata recovery and validation service. Bring actual read and write patterns to the trial discussion.
A business worth understanding.
Suggest your business or one you find interesting. Tell us what you want to understand about its product, positioning, design or workflows.
Suggestions are free. Selection and publication stay with the desk.
- lakeFS platformConsulted
- lakeFS data model and internalsConsulted
- lakeFS architectureConsulted
- Actions and validation hooksConsulted
- lakeFS Enterprise pricingConsulted
- DVC project acquisitionConsulted
- lakeFS community source repositoryConsulted

