DSS 15 Release notes

Migration notes

How to upgrade

Pay attention to the warnings described in Limitations and warnings.

Migration paths to DSS 15

Limitations and warnings

Automatic migration from previous versions is supported (see above). Please pay attention to the following cautions, removal and deprecation notices.

Cautions

LDAP TLS certificate hostname enforcement

Enhanced LDAP security: following an Apache library security fix, DSS now verifies that the hostname in an LDAP server’s TLS certificate matches the configured LDAP server hostname.

Spark 4 upgrade

For customers using Dataiku with Spark-on-Kubernetes (which includes all Dataiku Cloud and Cloud Stacks customers), Spark has been upgraded from 3.5 to 4.1.

Please carefully read Spark 4.0 and 4.1 release notes to be informed of potential changes to your Spark recipes.

In addition, Spark 4.1 dropped support for Python 3.9. Pyspark recipes and notebooks using Spark and still running Python 3.9 will fail. In order to facilitate this transition, DSS will automatically detect these code environments and prompt administrators to upgrade them to a more recent Python version at startup. Please follow the instructions.

This upgrade procedure is available both on Design and Automation nodes. Please however note that if you import an old bundle containing a Python 3.9 code env to an automation node after the upgrade, it will not be upgraded. We strongly recommend making sure to generate new bundles after upgrading the code envs of the design node if you need to update bundles on the automation node.

Upgrade of the builtin Python env

The builtin Python env is automatically updated to Python 3.11 or Python 3.12 (depending on your OS) upon upgrade. As a reminder, you cannot install any package in the builtin env, and it’s not usually recommended to use it for your recipes.

Furthermore, Flask is upgraded to 3.1 in the builtin environment. This Flask version removes some previously deprecated functions. Webapp backends using Flask on the builtin environment may need to be updated, or moved to a code environment using an older Flask version.

Air-gapped document extraction

DSS 15 uses a different hierarchy of model resources for structured extraction, requiring an update of the internal code environment if it already exists

After upgrading DSS:

  • If your DSS instance has Internet access: update that code environment with the “Rebuild env” option checked. The model resources will be automatically re-downloaded the next time they are needed.

  • If your DSS instance does not have Internet access, re-run the offline setup for structured text extraction. See Initial document extraction setup for more details.

Or (not recommended), if you prefer to re-use the models you already downloaded, an administrator with shell access must

  • Uncheck the “Rebuild env” and update the code environment; this will move the model resources to their new place

  • From a shell console, make a copy of the moved model resources (the hierarchy under <DATA_DIR>/code-envs/resources/python/INTERNAL_document_extraction_v1/document_extraction_models/)

  • Update the code environment again, with the “Rebuild env” option enabled: this clears the code environment (including the model resources) and updates its packages

  • From a shell console, restore the model resources

CUDA upgrade for non-containerized local LLM inference

If you are using local LLM inference without containers (this is a rare setup), you need to upgrade your CUDA runtime to version 13.

Support removals

Some features that were previously announced as deprecated are now removed or unsupported.

  • Summarization-only models in Hugging Face connections are not supported anymore. Text summarization can still be used with text generation LLMs.

  • Amazon Linux 2 is not supported anymore

  • SuSE 15 SP4 is not supported anymore

  • AlmaLinux 8 support for container images is not possible anymore. Please remove the --distrib flag.

  • New code envs cannot be created anymore with Python 3.6, 3.7 or 3.8

In addition, the following plugins have been removed and cannot be used anymore with DSS 15:

Deprecation notices

DSS 15 deprecates support for some features and versions. Support for these will be removed in a later release.

  • Support for code-envs using Python 3.9

  • Support for code-envs using Python 3.10

  • The Dataiku Code Assistant for Code Studios. Please instead use one of OpenAI Codex, Claude Code, OpenCode or GitHub Copilot

  • Support for using Conda to create and manage code envs

Version 15.0.0 - August 14th, 2026

DSS 15.0.0 is a major upgrade to DSS with major new features.

New feature: Agent Skills

Agent Skills are now native in Dataiku for building Visual Agents.

Skills allow you to package reusable Agent know-how & resources in a Skill that you can use in multiple Visual Agents. Skills benefit from progressive discovery, with the agent initially only receiving the skill description, and deciding autonomously when and whether to load it, and when or whether to load additional resources, keeping the context focused on the task at hand.

For more details, please see Agent Skills.

New feature: MCP server

Both Agents and Tools can now be exposed through Dataiku’s native MCP server, for easy consumption by external agentic systems.

New feature: Polars support

The Dataiku Python API now supports Polars, a popular and usually faster alternative to Pandas.

For more details, please see Datasets (reading and writing data)

New feature: Expanded Python fast-path support

Dataiku now supports fast-path reading in Python recipes for more data sources, and now also supports fast-path writing.

Fast-path reading as Pandas dataframes is supported for:

  • Snowflake datasets

  • S3 datasets using Parquet or CSV

  • Azure Blob Storage datasets using Parquet or CSV

  • GCS datasets using Parquet or CSV

Fast-path reading as Polars dataframes is supported for:

  • S3 datasets using Parquet or CSV

  • Azure Blob Storage datasets using Parquet or CSV

  • GCS datasets using Parquet or CSV

Fast-path writing of Pandas dataframes is supported for:

  • Snowflake datasets

  • S3 datasets using Parquet or CSV

  • Azure Blob Storage datasets using Parquet or CSV

  • GCS datasets using Parquet or CSV

  • BigQuery, Databricks, Redshift, Trino and Synapse datasets (if the fast-path writing option is enabled on them)

Fast-path writing of Polars dataframes is supported for:

  • S3 datasets using Parquet or CSV

  • Azure Blob Storage datasets using Parquet or CSV

  • GCS datasets using Parquet or CSV

  • Snowflake, BigQuery, Databricks, Redshift, Trino and Synapse datasets (if the fast-path writing option is enabled on them)

For more details, please see Datasets (reading and writing data)

New feature: Multi-target regression

Multi-target regression in Visual AutoML lets you train models that predict several numerical targets.

For more details, please see Multi-target Regression

New feature: On-the-fly automatic build of Code Env and Code Studio images

Previously, when upgrading DSS, container images for Code Envs and Code Studios needed to be rebuilt. Until they were, workloads using these Code Envs or Code Studios, including recipes and webapps, failed.

A new automatic build system has been introduced. The first time a workload requires a Code Env or Code Studio image after an upgrade, DSS will automatically build the image and start the workload once the image is built. First runs will be slow while the images build, but no workload will fail.

Agentic AI & RAG

  • New feature: Debug mode for Visual Agents lets you visually inspect the steps of your Agent while it runs your test queries

  • New feature: The OpenAPI Agent Tool lets you easily query OpenAPI endpoints from Agents, with automatic discovery of the endpoints and ability for the builder to choose which ones to expose

  • New feature: Knowledge Banks: Snowflake Cortex: added ability to create a Search Service from a Dataset, and obtain a Knowledge Bank that synchronizes with the dataset’s table on the Snowflake side

  • New feature: Knowledge Banks: support for Databricks AI Search as the vector store backing unmanaged Knowledge Banks

  • Structured Visual Agents: added support for human confirmation of the tool calls in Mandatory/Manual Tool Call and Delegate to Other Agent blocks

  • Structured Visual Agents: added ability to declare additional dependencies (e.g., used by Python blocks)

  • Structured Visual Agents: CEL expressions: added support for dict.has("key") and dict.get("key", "default value")

  • Structured Visual Agents: Python block: fixed setting a state variable from a Python block used as “Before”/pre-block

  • Structured Visual Agents: fixed deletion of a block used in an exit condition of another block

  • Simple Visual Agents: added support for human confirmation of the tool calls in the LLM Mesh Query tool

  • Agents: The new “Expose” screen shows to the builder all options for reusing an Agent built in Dataiku

  • Agents: Added ability to stop a running Agent, Agent Tool or Retrieval Augmented LLM

  • Agents: Performance improvements for high-throughput agents

  • A2A Server: now supports versions 0.3 and 1.0 of the A2A specification and better handles Agent-set state

  • A2A Server: increased timeout to 30 minutes

  • A2A Server: fixed a possible hang of the request and Agent

  • Agent Evaluation: Added a guided setup to configure agent interaction logging and create a preconfigured Agent Evaluation recipe and evaluation store directly from an agent. The setup is available to non-admin users with the required permissions.

  • Agent Evaluation: Added a conversation explorer for multi-turn evaluations, allowing users to inspect the full conversation, each turn’s trajectory, and the associated metrics in a single view. Human-in-the-loop interactions, including those from Structured Visual Agents, are grouped into the relevant logical turn.

  • Agent Evaluation: Improved diagnostics when conversation metrics are enabled but the input data contains no conversation identifiers.

  • Agent Review: The review details panel now shows the number of tests and runs, as well as the last run date.

  • Agent Review: Improved the experience for reviews without tests by displaying a dedicated empty state and opening the test creation view when appropriate.

  • Agent Review: Fixed keyboard navigation issues that could unexpectedly open the test panel or leave the wrong tab active.

  • Knowledge Banks: added ability to retrieve multiple columns, including embedded content, in multimodal KBs

  • Knowledge Banks: fixed indexing of string metadata with trailing zeroes

  • Knowledge Banks: fixed the link to the source Folder of a Knowledge Bank, when a KB Search tool or retrieval-augmented LLM is in a different project than the KB/Folder

  • Knowledge Banks: fixed a possible hang while reading the details of a Knowledge Bank that’s already in use elsewhere

  • Knowledge Banks: Milvus (remote): you can now configure the indexing type of the embedding field

  • Knowledge Banks: Azure AI Search: added support for metadata objects/subfields in unmanaged Knowledge Banks

  • Knowledge Banks: Azure AI Search: fixed embedding of large document batches

  • Knowledge Banks: Elasticsearch: fixed display of sources

  • Knowledge Banks: Milvus: added support for metadata objects/subfields in unmanaged Knowledge Banks

  • Knowledge Banks: OpenSearch: fixed the combined usage of custom headers and global proxy

  • Knowledge Banks: Pgvector: fixed a possible race condition when simultaneously building multiple Knowledge Banks on a PostgreSQL instance where the vector extension is not yet built

  • API: improved Python API to access artifacts and sources of a response

  • API: Agent Tools: added a method to use a DSSAgentTool object directly in a completion query

Document extraction

  • Extract Fields recipe: Added auto-detection/suggestion of the fields to extract

  • Extract Content recipe: Added ability to export tables from documents into CSV files in the recipe’s output folder

  • Extract Content recipe: fixed local files cleanup upon project deletion

  • Extract Content recipe: Raw Text extraction: extended the OCR-on-images option to images inside DOCX and PPTX documents

  • Extract Content recipe: Fixed offline usage of image classifier model when using structured extraction on non-PDF files

LLM Mesh

  • New feature: A new API for Persisted Conversations lets you query LLMs and Agents without needing to maintain the conversation history on the client side. The conversations are stored in a Dataiku-managed database, and you need only send the next user message.

  • New feature: Leverage the Web search tool in LLM calls. This can be used directly in Prompt Studio, Prompt Recipe, … and also when building Agents. Supported on OpenAI, Anthropic, Azure OpenAI and Microsoft Foundry connections.

  • New feature: Rate limiting based on tokens-per-minute limits

  • Anthropic / Bedrock / Foundry: added support for Claude Sonnet 5 and Fable 5 models

  • Bedrock: added support for the Nova 2 multimodal embedding model

  • Bedrock: added support for OpenAI models using the Mantle Responses API

  • Vertex AI: added multi-region locations for models that support it

  • Vertex AI: added options for custom URL & headers

  • Vertex AI: improved configuration of individual models

  • Microsoft Foundry: fixed chat completion with an explicit tool_choice

  • Local models: added support for DeepSeek V4 Flash

  • Local models: added support for inference of some models without a GPU

  • Local models: added support for multimodal output of Agent Tools

  • Local models: improved speed of transferring model weights from the model cache to containers

  • Local models: fixed multi-turn conversations on some Mistral models after a tool call

  • Local models: fixed abidance to max tokens setting on embedding models

  • Prompt Studios: a Prompt Studio created from a Prompt recipe now offers to update said recipe when Exporting as recipe

  • API: added ability to try asking an LLM for JSON even if it does not support structured output

  • API: fixed handling of additional context when calling Agents using LangChain wrappers without streaming

  • API: limited the growth of memory fragments in multi-turn conversations

  • API: OpenAI-compatible API: fixed placement of the status field

  • API: OpenAI-compatible API: fixed missing id field in non-streamed Chat Completion responses

  • MCP connection: custom headers marked as secret are now stored encrypted

  • Project import: fixed import failure when some LLM IDs are malformed

Dataiku AI & Cobuild

  • New feature: Per-user custom instructions

  • New feature: Cobuild can now create and manage ML evaluation recipes and Agent evaluation recipes

  • Cobuild does not open by default for users without Cobuild access to the project

  • Fixed search (using Cmd+F / Ctrl+F) within Cobuild conversation

  • Cobuild now automatically names conversations

  • Improved behavior of Cobuild on small screens when on the Flow

  • Added ability to pin conversations

  • Fixed possible lag of conversations on very large instances

  • Fixed minor display issues on Firefox

  • Added ability for administrators to block project-level custom prompts

  • Fixed failures due to too large context when training ML models

Machine Learning

  • New feature: Catboost support in Visual AutoML, with its native categorical feature handling.

  • Train recipe: added option to skip the computation of expensive reports

  • Model training: K-means clustering models: sped up the computation of feature importance

  • Model training: K-means clustering models: the “Number of tested initial centroids” can now be left empty (automatic)

  • Model training: ETS time series forecasting models: fixed relay of the error in some failed training situations

  • Model training: fixed failure when all values of a categorical feature are dates

  • Model scoring: added support for optimized scoring of models using ICA-based feature reduction

  • Model scoring: added support for optimized scoring of Isolation Forest, KMeans and mini-batch KMeans models

  • Model scoring: TFT & NHITS time series models: fixed scoring of a dataset that includes series identifiers that were unknown at train time

  • Model export: fixed exported Python scoring of Stochastic Gradient Descent classification models using Modified Huber loss

  • Model export: fixed possibly incorrect continuous value in generated optype attributes of exported PMML

  • Trained models: added a new API to export raw explanations and observations from feature importance

  • Trained models: fixed missing “Training data” information on some models (k-fold, time series)

  • Trained models: fixed slow loading of forecast report on time series forecasting models

MLOps

  • MLflow models: Fixed an issue where the model-version import dialog failed to load, preventing users from importing a first or subsequent version into a saved model.

  • Unified Monitoring: Fixed project filters and health distribution charts failing to display when some monitored projects have no deployment stage.

Dataset and Connections

  • New feature: Added support for ClickHouse

  • GCS: Fixed syncing to GCS when authenticating with a p12 private key file

  • Added automatic fast write support for Fabric Warehouse with Spark

  • Iceberg: Added ability to skip TLS certificate validation in REST catalogs

  • Iceberg: Added necessary libraries to connect to Hive catalogs

  • Iceberg: Fixed S3 remote signing on REST catalogs

  • Snowflake: Fixed issue with dates in the Snowflake->Cloud Storage unload

  • Databricks: New connections now default to version 3 of the JDBC driver. Note that some edge case behaviors may change. This version fixes issues with the “datetime without timezone” data type.

  • Databricks: Fixed fast-write on an Azure Blob connection using per-user OAuth credentials and a client secret

  • SFTP: Fixed creation of directories on servers that only support a single connection

  • FTP: Fixed possible failures on remote files that have no or bad timestamp

  • BigQuery: Improved error message when missing project id and/or dataset id

  • SCP: Fixed possible hang

  • GCS: Fixed observance of Private Service Connect endpoint when reading non-Parquet datasets

  • Snowflake: Added a new recipe engine option to specify the max file size when using the fast path to unload a Snowflake dataset to cloud

  • Snowflake: Respect the compression method param when using the fast path to unload a Snowflake dataset to cloud (for supported compression methods)

  • S3: Fixed display of selected bucket

  • SQL datasets: Fixed “Get tables list” after clearing the table field

  • MySQL: improved detection of timezone with older driver versions

  • Sharepoint: improved dataset creation from a single file in a Managed Folder

Flow

  • New feature: when dropping a file to the Flow, you can now choose to upload it in a Managed Folder

  • Improved display of dataset information in the “record count” view

  • Fixed possible failure with managed folders using non-standard identifiers

Recipes

  • Join: Improved performance for “send unmatched data” on DSS engine

  • Join: Fixed DSS engine wrongfully discarding leading/trailing spaces in join fields

  • Prepare: the data source for the “GeoIP” processor has been changed; some data may change as a result

  • Fixed possible inaccurate “Slow” indication on Spark engine label in recipes

  • Fixed inaccurate warning about “Computing execution plan in a transaction” when running recipes

Charts and Dashboards

  • New feature: Dashboards: synchronized drill-down within hierarchies across charts

  • Pivot: Added percentage scale compute mode for table with multiple columns or rows

  • Line charts: Fixed ordering of lines in legends

Dataiku Stories

  • Pivot tables: Measures as Rows or Columns: added a toggle input to switch between measures as columns or rows (default)

  • Pivot tables: Expand/Collapse: added option to expand / collapse rows & columns in both Edit mode (under “Chart options”) and View mode

  • Pivot tables: Total Display Toggles: split total display options into independent toggles for row and column totals

  • Pivot tables: Empty Values Handling: added options for managing empty values display

  • Pivot tables: Interactive Filtering: in view mode, table headers can now be clicked to filter the slide directly

  • Pivot tables: Style Options: added distinct styling for row headers, column headers and values

  • Pivot tables: Theming Properties: integrated table styling into the theming system for consistent customization

  • Pivot tables: Column Sizing: added options to auto-size all columns, and to resize columns to fit all data within the table view

  • Stacked bar charts: added ability to show or hide totals and define the font properties

  • Message on charts: improved the visibility and consistency of error messages

  • Numeric axis range settings: The numeric axis supports both automatic and manual range settings. By default, the automatic range is active. When manual range is selected, the “force inclusion of zero” option becomes disabled and it’s possible to manually define the range values for the axis. These settings are also supported for charts with dual axes.

  • Improved data zooming

  • Fixed “Add dataset” form failure when a connection is deleted

  • Fixed naming of new slides

Deployer

  • Fixed deployment failure when a group is missing on the automation node

Governance

  • New feature: Custom page designer: added a Grid system with multiple widgets.

  • New feature: Custom page designer: added ability to natively define aggregating visualizations. Visualizations support synchronized drill-down between charts on the grid.

  • New feature: New configuration options for tab content and organization.

  • Fixed email settings wrongfully appearing as disabled

  • Removed the support of the previously deprecated mayManageGovern key from the public API (global permissions) that was renamed to isGovernArchitect in 14.4.0.

  • New feature: Blueprint Versions: post-phase hooks are executed right after the item’s action is fully committed into the database. Useful to run custom python code requiring a confirmed, successful action, or long-running script (e.g., calling a third-party service).

  • Added settings in text field view components to choose its default display(markdown or plain text) and editor (rich-text, single-, or multi-line raw text).

  • Added metadata in public API: item’s creation date & user, last modification date & user.

  • Fixed display of some read-only fields.

  • Fixed status of the sign-off review button when an item is being edited.

  • Removed the effective governance settings section for items that are already governed.

Dataiku Applications

  • New feature: added “change input dataset” tile

  • Added ability to directly display a dashboard within the Dataiku Application

  • App as recipes: Faster write to the output datasets through usage of native sync capabilities

  • Fixed “Manage tags” button

Workspaces

  • Fixed display of global tags in the mass action set tags of the objects in the workspace

  • Fixed display of customized color for tags on the consumer homepage

  • Fixed display of tags for links in card display mode

  • Workspace user search now works both by login and display name

  • Improved display of user icons on workspace cards

  • Improved display of workspace card style in the left panel workspace list

  • Better display when user does not have permission to view a workspace

Collaboration

  • The Data Catalog is now the Catalog and includes more asset types, notably models and agents

  • Enterprise Asset Library: Added APIs to export/import assets

  • Consumer home page: Fixed custom tag color display

  • Consumer home page: Displayed tags in list view

  • Consumer home page: Fixed promoted content custom thumbnail

  • Consumer home page: Fixed switching between list and tile views with filters active

  • Version Control: Added ability to set the Git “committer” in addition to “author”

  • Version Control: Fixed possible malformed Git commands when setting config values

  • Version Control: Improved per-user-SSH-key UI

  • Version Control: Fixed possible tag conflict when using remote Git repositories together with projects being bundled

  • Version Control: Fixed possible failure when reading status on duplicated projects

  • Project name is now included in browser tab titles

  • GDPR plugin: display connection description

  • Added code samples for Python probes and SQL probes

  • Added sections to the plugin store UI

  • Fixed items that could still appear in “Recent & Favorites” after being deleted

  • Fixed refresh of dashboard thumbnails

Scenarios and Automation

  • Added an “end date” option for scenario time based triggers

  • Improved “last run” tab refresh after triggering a manual run of a scenario

Coding & API

  • Added support for Pandas 3

  • Improved error when running a code recipe that relies on a shared project lib for which the user does not have proper permissions

  • Fixed the “read dataset row by row” starter notebook code

  • Webapps: Fixed dataiku.WebappImpersonationContext with FastAPI webapps

  • Fixed hard-to-read project libraries code when user does not have write access

  • Fixed possible race condition with the DSSScenarioRun.refresh method

  • Added API to create Code Studio templates

  • Added support for plugin datasets in autodetect_settings

Code Studios

  • Fixed wrongful warning about unsynced changes when stopping a VSCode Code Studio

  • Fixed unwanted start when editing a stopped Code Studio Webapp

  • Fixed deployment of coding agents when Code Studio uses dynamic Kubernetes namespace resolution

  • Fixed deletion of the Kubernetes deployment when deleting a Code Studio

  • Deprecated the “Install Copilot” checkbox – Copilot is now builtin in VSCode

  • Codex: Fixed issue with numerical logins starting with 0

Notebooks

  • Prevented renaming a code notebook with special characters

  • SQL notebook: Improved support of variable expansion

  • SQL Notebook: Fixed scatter plot chart

Elastic AI

  • EKS: Fixed possible cluster autoscaling failures caused by missing region information

  • EKS: Fixed autoscaling failure when using “latest” as cluster version

  • EKS: Fixed possible Metrics server failure with recent eksctl versions

  • EKS: Only run the nvidia daemonset on GPU nodes

  • Fixed missing progress deadline for webapp deployments on Kubernetes

  • Pods now set runAsNonRoot by default

  • Fixed execution of “Export notebook” step when missing properly-built Spark images

  • Fixed possible race condition when killing containers while they are starting

Cloud Stacks

  • AWS: Fixed reprovision of EC2 instances with termination protection enabled

  • Azure: Fixed private DNS zones support

Security

Performance & Scalability

  • VSCode settings are now stored in a non-versioned folder, to avoid possible instance lags while syncing very large user settings

  • Fixed possible crash when exporting very large SQL notebook query result to Excel

  • Cgroups: When enabling cgroups, cgroups V2 is now the default

  • Cgroups V1: Fixed JEK CPU not properly taken into account

Misc

  • Added support for RedHat 10, AlmaLinux 10, RockyLinux 10, OracleLinux 10

  • Python 3.14 support is no longer experimental

  • Data Lineage: Fixed mass action in manual remapping modal

  • Fixed issue building internal code env images in the API

  • Fixed lag when dismissing the “How likely are you to recommend Dataiku?” popup

  • Improved confirmation message when deleting multiple connections

  • Fixed the default Todo list links

  • Added safeties against accidental downgrades of the runtime database schemas

  • Fixed issue with custom format export plugins

  • Fixed possible failure of project creation macros

  • Fixed DATE type display in plugins

  • Explore: improved scrolling in filters

  • Formula: Fixed issue with inc operator on dates