kedro

Major features and improvements

• Added an optional params query parameter to GET /snapshot for resolving ${runtime_params:...} interpolation per request, using the same format as kedro run --params.
• Graduated llm_context_node, LLMContextNode, LLMContext and tool from experimental to stable. They no longer emit a KedroExperimentalWarning.

Breaking changes to the API

• Security fix: Custom logging handler, filter or formatter classes referenced in conf/logging.yml are no longer imported unless their module is on the logging allowlist. Only the logging standard-library package (including logging.handlers) and kedro.logging are trusted by default. Projects that point conf/logging.yml at their own logging classes must now list those modules in the KEDRO_LOGGING_MODULE_ALLOWLIST environment variable (comma-separated) to keep them working.

Major features and improvements

• Added dataset validation: datasets can declare a validator in their catalog entry, and the DataCatalog enforces it on every load and save.
  • Failures raise DataValidationError (a DatasetError subclass), reporting every failed check at once.
  • Pandera is the built-in adapter, installed with the new kedro[pandera-pandas] or kedro[pandera-polars] extras; custom validator classes and functions also work.
  • The DATASET_VALIDATION setting and the KEDRO_DATASET_VALIDATION environment variable switch validation off.
  • validate_dataset and validate_catalog validate on demand, reporting structured ValidationResult outcomes instead of raising.
• Added --runner-params to kedro run, allowing runner constructor keyword arguments such as max_workers to be passed from the CLI.
• Added the node function name to project inspection snapshots as NodeSnapshot.func_name.
• Added node source location metadata (NodeSnapshot.source) to inspection snapshots for displaying node code.
• Added optional runtime_params to get_project_snapshot().

Bug fixes and other changes

• Fixed a thread-safety issue in _ProjectPipelines where concurrent access could trigger duplicate pipeline loading or operate on a cleared cache; all shared state mutations are now protected by a reentrant lock.
• Fixed a thread-safety issue in the Kedro HTTP server by adding a serving_mode argument to KedroServiceSession.create() that preloads all pipelines upfront.
• Fixed a RecursionError when initialising a session with dynaconf-backed settings by converting settings.SESSION_STORE_ARGS to a plain dict before deepcopying it.
• Excluded kedro_benchmarks from the built wheel so benchmark tests are no longer shipped with the package.
• Fixed get_close_matches returning duplicate suggestions when several inputs matched the same target, and being able to exhaust a one-shot iterable passed as targets.
• Fixed docs generated with kedro new --tools=docs so the Sphinx HTML build runs sphinx-apidoc and creates API docs.
• Deprecated --async flag for kedro run in favour of --runner-params=is_async=True.
• Added a warning in ParallelRunner catalog validation that identifies cloud-backed datasets (e.g. S3, GCS, Azure Data Lake) failing to pickle and suggests using ThreadRunner or SequentialRunner instead.
• Redacted credentials, signed-URL query parameters and fragments from dataset filepaths in AbstractDataset.__repr__, DatasetError messages, and HTTP server /snapshot and /run responses.

Documentation changes

• Added a documentation page for dataset validation.
• Documented the HTTP server's restriction on runtime_params-resolved catalog type fields, in the templating guide and the HTTP server guide.
• Added a new "Vibe coding with skills" page documenting the kedro-skills plugin.
• Added mdformat as a Markdown autoformatter, run via make lint and as a pre-commit hook, and reformatted the existing Markdown. Fixed several step-by-step guides where numbered lists were rendering as repeated "1." instead of counting up, because their content wasn't indented under the right list item.
• Fixed the DaskRunner example on the Dask deployment page.
• Fixed the open_args_load examples in the data catalog documentation to use text mode so the encoding option applies.
• Added a cheatsheet with Kedro concepts and CLI commands to the documentation.

Breaking changes to the API

• Security fix: A catalog type field resolved with runtime_params supplied through the HTTP server's POST /run no longer accepts an AbstractDataset class chosen by the request. runtime_params-driven type selection through other, trusted callers (for example, kedro run --params) is unaffected.
• The settings.py entry RUNNER_MODULES_WHITELIST has been renamed to RUNNER_MODULE_ALLOWLIST, following the <COMPONENT>_ALLOWLIST naming convention for Kedro al

Major features and improvements

• Added a new HTTP Server layer which uses the KedroServiceSession to execute pipelines from HTTP requests. It offers the following endpoints:
  • /run to run pipelines with optional runtime parameters.
  • /health to check the server's health status.
  • /snapshot to retrieve the current project snapshot.
• Added a new CLI command kedro server start to run the server.
• Added selective pipeline loading: when --pipelines is specified, only the requested pipeline modules are imported.

Bug fixes and other changes

• Fixed Rich logging integration so node input/output brackets render correctly in console logs and dataset colour markup does not leak into plain log handlers.
• Improved the AbstractDataset.from_config() error message for custom dataset classes that are still abstract, so it no longer suggests invalid constructor arguments when required dataset methods are missing.
• Fixed kedro new accepting project names whose derived package name shadows a Python standard library module or is a Python keyword (e.g. email, json, import), which silently produced a broken, unimportable project. Such names are now rejected at creation time with a clear message.
• Fixed kedro pipeline create accepting Python keywords (e.g. for, import, return) as pipeline names. Such names are now rejected at creation time with a clear error message.

Documentation changes

• Documented hooks limitation when using ParallelRunner.
• Updated the Amazon EMR Serverless deployment guide with a strategy-first layout, custom container images, Spark catalog configuration, and troubleshooting for common deployment issues.
• Updated the AWS Step Functions deployment guide with a strategy-first layout, pipeline-level namespace grouping, Lambda container deployment, and an end-to-end Spaceflights walkthrough.
• Updated the AWS Batch deployment guide with a strategy-first layout, custom AWSBatchRunner usage, namespace-based job submission, and S3-backed catalog configuration.

Community contributions

Many thanks to the following Kedroids for contributing PRs to this release:

• Feng Jikui
• Rudra Dudhat
• Jean-Baptiste Braun
• Gargi
• Simon

Major features and improvements

• Add official support for Python 3.14.
• Added inspection API to get project snapshot.
• Improved %load_node to include same-module helper dependencies via AST extraction, with explicit fallback warnings when extraction degrades to function-only source loading.
• Scoped parameter type-hint validation to the pipelines being run. When kedro run --pipelines=<name> runs one or more named pipelines, only those pipelines' node signatures are inspected for type hints, avoiding unnecessary work and spurious "conflicting type" warnings from pipelines that aren't executing.

Experimental features

• Added the first iteration of the KedroServiceSession, a new session implementation that allows for multiple runs and data injection. NOTE: This session implementation is under active development and may occasionally contain bugs or breaking changes. We encourage users to try it out and share their feedback with us.

Bug fixes and other changes

• Fixed parameter validation for optional typed params so explicit None values are preserved and do not trigger validation errors.
• Added review-kedro-pr agent skill (Cursor, GitHub Copilot) for Kedro-aware PR review with optional GitHub comment posting.
• Removed outdated TRANSCODING_SEPARATOR alias from kedro.pipeline.pipeline.
• Added an optional pydantic dependency extra, allowing users to enable Pydantic support with pip install "kedro[pydantic]".
• Fixed parameter validation for namespaced pipelines.
• Fixed runners iterating over iterable-but-non-generator node outputs (e.g. mne.Epochs) before saving them. Streaming behaviour is now restricted to generator-function nodes, so custom business objects are passed to catalog.save unchanged.

Documentation changes

• Added documentation for KedroServiceSession.
• Updated the API documentation for kedro.framework.session to include the new KedroServiceSession and AbstractSession classes.
• Added security model documentation covering trust boundaries, user responsibilities, and framework vulnerabilities.
• Updated advanced_configuration.md with custom loader info.

Community contributions

• BhavayChopra
• jeevan6996
• Rahul Bansod
• PragnyaKhandelwal
• datascienceio
• Guillaume Tauzin
• Prince Kumar