# v2.0.0

## Release Changelog

### v2.0.5

This release rolls up hotfixes on top of v2.0.4. There are no breaking changes or deprecations.

#### Release Availability Date

21-July-2026

**Recommended Versions**

- **CLI/SDK**: 1.6.0.4
- **Remote Executor**: v2.0.5-cloud, v2.0.4-cloud, v2.0.3-cloud, v2.0.2-cloud, v2.0.1-cloud, v1.1.3-cloud, v1.0.3-cloud
- **On-Prem Versions**:
  - **Helm**: 1.6.189
  - **API Gateway**: 0.7.3

#### Bug Fixes

- Fixed Search V2.5 returning duplicate entities within a page while dropping an equal number of other matches across pagination boundaries, so paginated search results are now complete and consistent.
- Fixed proposal listings (for example on Data Products) failing to load with an "all shards failed" error for users belonging to a large number of groups or roles.

### v2.0.4

This release rolls up hotfixes on top of v2.0.3.

#### Release Availability Date

17-July-2026

**Recommended Versions**

- **CLI/SDK**: 1.6.0.4
- **Remote Executor**: v2.0.4-cloud, v2.0.3-cloud, v2.0.2-cloud, v2.0.1-cloud, v1.1.3-cloud, v1.0.3-cloud
- **On-Prem Versions**:
  - **Helm**: 1.6.189
  - **API Gateway**: 0.7.3

#### Platform

- **ZDU stability**: Semantic search un-aliased indices are no longer deleted during orphan cleanup, preventing accidental index loss during zero-downtime upgrades.

#### Bug Fixes

- Fixed hidden documents incorrectly inflating search UI entity counts causing platform counts to appear much larger than what they are in reality.
- Fixed metadata propagation not matching target columns by simple field name, causing propagation to silently skip columns with nested field paths.
- Fixed inability to add multiple datasets to an existing incident entity via the API.
- Fixed ES8 kNN semantic search queries failing due to un-normalized `adjust_pure_negative` parameter.
- Fixed Search V2.5 rescore phase duplicating or dropping results.

### v2.0.3

This release rolls up hotfixes on top of v2.0.2.

#### Release Availability Date

07-July-2026

**Recommended Versions**

- **CLI/SDK**: 1.6.0.4
- **Remote Executor**: v2.0.3-cloud, v2.0.2-cloud, v2.0.1-cloud, v1.1.3-cloud, v1.0.3-cloud
- **On-Prem Versions**:
  - **Helm**: 1.6.189
  - **API Gateway**: 0.7.3

#### Bug Fixes

- Fixed an issue with resolving Slack integration connections, which caused the UI falsely indicate that Slack was not connected.

### v2.0.2

This release rolls up hotfixes on top of v2.0.1.

#### Release Availability Date

06-July-2026

**Recommended Versions**

- **CLI/SDK**: 1.6.0.4
- **Remote Executor**: v2.0.2-cloud, v2.0.1-cloud, v1.1.3-cloud, v1.0.3-cloud
- **On-Prem Versions**:
  - **Helm**: 1.6.189
  - **API Gateway**: 0.7.3

## Platform

- Stability improvements for ZDU upgrade when doing upgrades to `2.0.x` version.

### v2.0.1

This release rolls up hotfixes on top of v2.0.0.

#### Release Availability Date

02-July-2026

**Recommended Versions**

- **CLI/SDK**: 1.6.0.4
- **Remote Executor**: v2.0.1-cloud, v1.1.3-cloud, v1.0.3-cloud
- **On-Prem Versions**:
  - **Helm**: 1.6.173
  - **API Gateway**: 0.7.3

#### Product

- **Change History sidebar and documentation diffs** — a clock-icon button on the entity profile version picker opens a timeline drawer showing documentation, tag, ownership, schema, and version-milestone changes for the current or all versions of an entity. Documentation changes render as collapsible diffs.
- **Timeline API for versioned entities** — the timeline API can now return the change history of a versioned entity (currently Glossary Term) scoped to a single version or unified across the entire version set.

#### AI / Ask DataHub

- **SQL-context MCP tools share a single rollout flag.** `find_sql_context` and `draft_sql_for_tables` are now gated behind one `ENABLE_SQL_CONTEXT_TOOLS` flag (default off) so they roll out together; `generate_sql_sketch` keeps its own separate flag.

#### Platform

- **Actor-scoped policy authorization.** Session-scoped actor identity now flows through authorization checks, and resource resolution for a request is cached and resolved once per authorization request instead of repeatedly.
- **Faster semantic-search bridge-document backfill.** The backfill now writes a single full-document index update per bridge URN instead of per-aspect writes, batched with bounded parallelism — large-scale dataset semantic-search backfills now complete instead of stalling at production scale.

#### Breaking Changes

- **Linking a physical dataset to a logical parent now requires Edit Entity on both datasets** (datahub-project/datahub#17904). Previously enforcement only checked the child dataset. Creating or changing a `logicalParent` link now also requires **Edit Entity** on the proposed parent dataset; clearing a logical parent still only requires **Edit Entity** on the child. **Action:** grant **Edit Entity** on logical-parent datasets to principals who create or change these links, or update policies so those operations continue to succeed.

#### Security / Dependencies

- **Netty**: bumped to 4.2.15.Final to address CVE-2026-47691.
- **Jackson databind**: bumped to 2.21.4 to address CVE-2026-54512 / CVE-2026-54513.
- **Apache Shiro**: bumped to 2.2.1 to address CVE-2026-49268.
- **PyJWT**: bumped to 2.13.0 to address GHSA-xgmm-8j9v-c9wx.
- **redshift-connector**: bumped to >=2.1.14 to address CVE-2026-8838.
- Routine dependency refresh: spring-kafka, jline, cryptography, pyOpenSSL.

#### Bug Fixes

- Fixed a prompt-caching regression from v2.0.0 where Ask DataHub's system prompt included a sub-hour timestamp, invalidating the model provider's cache on every turn.
- Fixed mTLS certificates not reliably reloading on rotation.
- Fixed the Settings → Roles page crashing when a role's policy list contained a non-policy entity.
- Restored workflow form submission on deployments still running v1 Action Workflows.
- Fixed the "(0)" count incorrectly showing on the More filters dropdown when no filters are selected.
- Fixed policies using a `NOT_EQUALS` condition returning an error when view-based authorization is enabled.
- Fixed nested column stats not appearing for BigQuery datasets with lowercased field paths.
- Fixed incorrect links in translated (i18n) UI text.
- Allowed `EDIT_ENTITY_TAGS` to be used for tag writes via the REST API.
- Restored the homepage's MCP page templates to their pre-v2.0.0 layout after a regression.
- Restored DELTA-bounds anchoring for smart-assertion evaluation and fixed executor timeseries queries sending invalid time-range filters.

#### Environment Variables

- `ENABLE_SQL_CONTEXT_TOOLS` (default `false`): replaces the separate `FIND_SQL_CONTEXT_TOOL` / `DRAFT_SQL_FOR_TABLES_TOOL` gates — enables `find_sql_context` and `draft_sql_for_tables` together. `generate_sql_sketch` keeps its own `GENERATE_SQL_SKETCH_TOOL` gate, unaffected.

### v2.0.0

#### Release Availability Date

25-June-2026

#### Recommended Versions

- **CLI/SDK**: 1.6.0.4
- **Remote Executor**: v2.0.0-cloud, v1.1.3-cloud, v1.0.3-cloud
- **On-Prem Versions**:
  - **Helm**: 1.6.173
  - **API Gateway**: 0.7.3

##### Tested Remote Executor versions

DataHub Cloud v2.0.0 has been internally tested with the following Remote Executor versions:

| Remote Executor version | Status | Notes                        |
| ----------------------- | ------ | ---------------------------- |
| v2.0.0-cloud            | Tested | Recommended.                 |
| v1.1.3-cloud            | Tested | Supported with this release. |
| v1.0.3-cloud            | Tested | Supported with this release. |

#### New Feature Highlights

- **Search V2.5 is now the default** — cross-entity ranking with name-match boosting and diversity promotion, plus latency optimizations for query understanding and multi-match.
- **Bridge-backed semantic search** now covers datasets, charts, dashboards, glossary terms, and data products via hidden bridge documents in the document semantic index.
- **Action Workflows v2** — build multi-step approval workflows for any asset, not just access requests: route each step to the right approvers automatically (such as a dataset's owners or domain stewards), require single- or multi-person sign-off, show or skip steps and form fields based on the asset, and let requesters cancel or admins override a request mid-flight.
- **MCP Audit** — new audit surface for MCP tool invocations: GraphQL `mcpAudit` resolver, dedicated audit tab in Settings → AI, KPI charts and history table.
- **CDE Steward agent** — new 10-star governance agent for Critical Data Element compliance and certification workflows.
- **Internationalization (i18n)** — first-class infrastructure (feature flag + user settings), end-to-end string extraction across the application (entity tabs, search, settings, governance, ingestion source metadata, etc.), and initial DE translations.
- **Distributed rate limiting** across REST, GraphQL, and OpenAPI with per-endpoint token-bucket controls and configurable jitter.
- **Domain propagation** — automatic propagation of domain assignment across lineage and containment relationships, with attribution.
- **`note_metadata_observation` MCP tool** — replaces `register_feedback` and `note_sql_anchor_observation`, and raises ANNOTATE or POST_ATTACHMENT proposals so agent-recorded observations land in the Context Hub inbox for SME review.
- **Assertion SQL Query Replay** - You can now see the exact queries issued by assertions against your warehouse, enabling you to replay failures to find failing records.
- **Anomaly Detection Assertion Improvements** — Volume assertions that use Anomaly Detection will now see more accurate predictions, reducing false alarms. This improvement is most pronounced on datasets with large row counts. This improvement requires upgrading your remote executor.
- **`SecretService` caller guard** — non-system actors (including PATs with `MANAGE_SECRETS`) can no longer decrypt secrets; hardens the credential-access surface.
- **New ingestion sources**: ThoughtSpot, TimescaleDB, Airbyte; production-ready SAP HANA with calc-view lineage, stored procedures, and query usage.

- All changes in [https://github.com/datahub-project/datahub/releases/tag/v1.6.0](https://github.com/datahub-project/datahub/releases/tag/v1.6.0)
  - Note Breaking Changes: [https://datahubproject.io/docs/how/updating-datahub/](https://datahubproject.io/docs/how/updating-datahub/)

#### Product

- **Action Workflows v2** — a redesigned authoring experience for governance workflows beyond access requests: conditional steps and form fields, context-aware dropdowns populated from the catalog, dynamic approver routing, and quorum-based sign-off. Available in Private Beta — contact the DataHub Cloud team to enable it. See [Action Workflows](../workflows/action-workflows.md) for what's new, the [Workflow Tutorial](../workflows/action-workflows-tutorial.md) for a worked example, and the [Workflow Reference](../workflows/action-workflows-reference.md) for the full JSON syntax.
- **Stable, caller-supplied request IDs.** The workflow form-request API now accepts an optional `id`, producing a predictable request URN (`urn:li:actionRequest:<id>`) instead of a server-generated UUID — so external automation can deep-link to or fetch a request directly by its URN instead of searching for it after submission.
- **Context Generation Settings** — domain is now required for enabling Context Generation; end-to-end backend wiring through the curator agent. Tracking events added for context hub features.
- **Internationalization** — i18n infrastructure with feature flag and user-settings backend; first-class string extraction across entity tabs (schema, queries, observe, validations, incidents, documentation, summary), search, settings, governance (domains, glossary, structured properties), identity / permissions, ingestion source metadata, home pages, and shared components. Initial DE translation for settings pages.
- **UI redesign: nav and home hero** — fully collapsed nav redesign and home hero toggler.
- **External document UI improvements** — inline preview, read-only fields, last-synced indicator.
- **Document anchor pattern UX** — anchor patterns flattened into per-metric cards; sort and cap anchor patterns; inline editing of anchor pattern rendered text.
- **Glossary** — Redesigned cards and sidebar with semantic tokens. Added support for custom relationships between glossary terms: create a Structured Property and select "Treat as Relationship" to define a new relationship type. Then navigate to the Related Terms tab on any glossary term and choose your custom relationship type when adding a related term.
- **Schema field drawer** — column statistics shown in full inline on the About tab.
- **Dataset Summary** — Generate Documentation flow now reachable from the Dataset Summary page.
- **Inbox auto-redirect** — Task Center inbox auto-redirects to proposals when the tasks tab is empty.
- **Tracking events** — context hub feature usage tracked for product analytics.

#### AI / Ask DataHub

- **MCP Audit** — new GraphQL `mcpAudit` resolver and Settings → AI → MCP Audit tab with feature-flag gate. KPI charts, history table, session and event drawers. `MCP_TELEMETRY_CAPTURE_PAYLOADS` env flag controls payload capture. `McpServerRequest` analytics report script.
- **MCP Apps infrastructure** — build infrastructure for in-product MCP App surfaces; `mcp/` ↔ `mcp_integration/` boundary enforcement for OSS hygiene.
- **`note_metadata_observation` MCP tool.** Replaces `register_feedback` and `note_sql_anchor_observation` with a single tool covering both metadata gaps and SQL-anchor quality observations. Update agent prompts that reference the old tool names.
- **`note_metadata_observation` raises ANNOTATE proposals on context docs.** When called with one or more `urn:li:document:` URNs in `related_objects`, the tool raises one `DOCUMENTS_PROPOSAL` ActionRequest per target with `proposalType: ANNOTATE`. The proposal stages a draft Document carrying an `ENTITY_ANNOUNCEMENT` Post; until an SME accepts it from the Context Hub inbox the target doc is unchanged. On acceptance, only the Post migrates to the target.
- **`note_metadata_observation` raises POST_ATTACHMENT_PROPOSAL when the agent cannot pin the gap to an existing doc.** Emits two coordinated MCPs: a Post with `postType=AI_OBSERVATION` and `target=null`, plus a new `POST_ATTACHMENT_PROPOSAL` ActionRequest. The Post is invisible across existing UI surfaces by construction; the ActionRequest shows up in the Task Center for SME triage. On accept, the Post becomes a real Comment on the chosen entity.
- **`DocumentProposalService.applyDraftDocumentChange` respects `proposalType: ANNOTATE`.** ANNOTATE proposals now skip the info / properties copy and only migrate Posts; EDIT, STATE_CHANGE, and CONFLICT acceptance behavior is unchanged.
- **`preview_sql_context` / `save_sql_context` use a clean preview-then-commit flow.** `preview_sql_context` builds the MCP App render payload without writing anything; `save_sql_context` commits on Approve in a single write. Cancel is handled client-side — no tombstone cleanup needed. Draft-based save flow with MCP App preview.
- **`find_sql_context` improvements** — consumes admin-curated `overrideSql`; deprecates `generate_sql_sketch`. Looker view-text and dbt model-text fallbacks added. SQL-override prop supported on semantic-anchor docs for human edits / proposals.
- **Semantic-anchor enrichment** — dataset bridge documents now include structured properties, field terms/tags, doc links, and domains. Dialect detection and Bedrock retry resilience.
- **Per-pattern question generation** — semantic anchors generate distinct questions per pattern; metricKeys lookup property for metric anchors.
- **CDE Steward agent** — new 10-star governance agent for Critical Data Element compliance + certification.
- **Agent authoring foundation** — 10-star authoring foundation with curator Cloudsmith wiring and runtime env-flag gating for SYSTEM agents. `Context Drop` and `Context Curator` agents default to off.
- **Ingestion-agent tooling** — Read/Grep/Glob source-browsing tools for connector troubleshooting from inside the agent.
- **LLM telemetry and billing** — `LLMCallEvent` per-call telemetry (token-billing primitive), context-local cost accumulation, per-turn usage logging, surface and time-to-first-token captured. New `POST /openapi/v1/billing/usage` endpoint.
- **Slack bot service-account mappings.** Admins can map a Slack `bot_id` to a DataHub corp user URN under Settings → Platform → AI → Enable Ask DataHub in Slack. When a registered bot @-mentions DataHub, the question is attributed to the mapped service account. Useful for cop-rotation bots, ticketing bots, and other automations.
- **Smart search** — surface `externalUrl` and inject DataHub URLs into results.
- **Confluence connector — HTML → Markdown** body conversion via `markdownify`.

#### Observe

- **Anomaly Detection Assertion Improvements** — Volume assertions that use Anomaly Detection will now see more accurate predictions, reducing false alarms. This improvement is most pronounced on datasets with large row counts. This improvement requires upgrading your remote executor.
- **Assertion ownership** — assertions now support owners. Ownership is viewable on both the Data health dashboard and the dataset quality tab. Monitoring rules will reflect the rule owner as the generated assertion owner.
- **Assertion SQL Query Replay** - You can now see the exact queries issued by assertions against your warehouse, enabling you to replay failures to find failing records.
- **Custom SQL assertions can optionally allow stored-procedure `CALL` statements.** `DATAHUB_EXECUTOR_ALLOW_CALL_STATEMENTS=true` lets a custom SQL assertion's statement be a `CALL my_db.my_schema.my_proc()` in addition to the read-only query shapes always allowed. Off by default; enabling accepts the risk that the procedure may perform mutations.
- **Assertion failure severity configuration SDK** — programmatic configuration of assertion failure severity behavior.
- **Severity-escalation broadcast in Slack** - changes in failure severity will be broadcasted to the main thread in Slack notifications.

#### Platform

- **`SecretService` caller guard.** Non-system actors (including human users and PATs with `MANAGE_SECRETS`) can no longer decrypt secrets. Controlled by `SECRET_SERVICE_CALLER_GUARD_MODE` (`ENFORCE` / `AUDIT` / `DISABLED`). Components that fetch secrets at runtime use system credentials in standard deployments and are unaffected.
- **Distributed rate limiting** across REST, GraphQL, and OpenAPI with per-endpoint token-bucket controls and configurable jitter (`RATE_LIMITS_RETRY_AFTER_JITTER_PERCENT`).
- **RFC 8693 token exchange** for trusted external issuers (OAuth2 token-exchange grant type).
- **Agent lifecycle stages.** Disabled SYSTEM agents are now `ARCHIVED` via `Status.lifecycleStage` instead of disappearing from the API. A startup reconciler reads `agent-flags.yaml` and restores the previous non-ARCHIVED stage when the flag flips back on.
- **Domain propagation** — automatic propagation of domain assignment across lineage / containment relationships, with attribution.
- **Bridge-backed semantic search** for datasets, charts, dashboards, glossary terms, and data products. Deploy GMS and MAE consumer together so source-entity metadata changes create, update, and delete bridge documents consistently. After enabling for a new entity type, run the bridge-document backfill and generate embeddings.
- **Search V2.5 cross-entity ranking** — DisMax scoring, name-match signals, diversity promotion across query understanding, multi-match, and focused fields. kNN inner_hits chunk text surfaced for the Cohere reranker.
- **Elasticsearch 8.18+ semantic search** — DataHub Cloud now supports semantic search on Elasticsearch 8.18+ deployments alongside OpenSearch. GCP deployments use a managed Vertex AI embedding provider (`gemini-embedding-001`); AWS deployments continue to use AWS Bedrock with Cohere Embed v3.
- **Per-entity OpenSearch / Elasticsearch mapping ceilings.** New `ELASTICSEARCH_INDEX_ENTITYMAPPINGLIMITS_<ENTITY>_<LIMIT>` env vars configure per-entity mapping limits; the configured value is baked into the index settings at creation/reindex time and pushed to existing live indices on every system update run. Use `DEFAULT` as the entity name for a fallback.
- **Event filtering framework with pre-deserialization MCL optimization** — drops uninteresting MCL events before deserialization, reducing GMS/MAE consumer load.
- **OTel GraphQL operation tracing** — OpenTelemetry instrumentation across GraphQL resolvers.
- **`RelationshipChange` platform event** — emitted on relationship changes for downstream consumers.
- **Domain attribution** — domain assignment now records the source of the assignment (manual vs. propagated) for audit purposes.

#### Ingestion

**New ingestion sources:**

- **ThoughtSpot** — new BI source connector.
- **TimescaleDB** — new connector supporting self-managed and Tiger Cloud TimescaleDB.
- **Airbyte** — new connector for Airbyte metadata.
- **SAP HANA** — production-ready connector with calc-view lineage, stored procedures, and query usage.

**Connector improvements:**

- **Hex** — Major in-place upgrade: upstream lineage (table-level and column-level), Project → Component links, run history (`lastRefreshed`), and optional AI context documents extracted directly from Hex REST APIs. New `include_lineage`, `use_queried_tables_lineage`, `connection_platform_map`, and `include_context_documents` config options. Hex Components are now ingested as Chart entities (see Breaking Changes).
- **Snowflake** — Internal Marketplace support; dynamic-table lineage extracted from `DYNAMIC_TABLE_GRAPH_HISTORY`; private-link Snowsight base URL override; Sweden Central Azure region mapping; fix for silently-dropped views in batched `SHOW VIEWS`.
- **Databricks Unity Catalog** — extract primary key, foreign key, and partition key constraints; opt-in Metric View ingestion; v1.1 composable lineage with agent metadata; default profiling switched to SQLAlchemy (from Great Expectations); ownership and `datasetProperties` emitted as standard MCPWs with incremental config; partner/DataHub user-agent for Databricks telemetry.
- **BigQuery** — Workload Identity Federation (WIF) auth.
- **Tableau** — support for virtual connections; Initial SQL ingested as lineage and custom property.
- **Glue** — PATCH mode for dataset properties; column-level Lake Formation tags by default (`extract_lakeformation_column_tags`); optional propagation of database tags to tables and columns (`propagate_lakeformation_tags`); inherited tags marked with propagation attribution.
- **Dremio** — query lineage and view-parent lineage respect `schema_pattern` and `dataset_pattern` (and skip the `_accelerator_` reflection schema); platform mappings for `BIGQUERY`, `RESTCATALOG` (Polaris OSS, Nessie, AWS Glue Iceberg REST, S3 Tables, Confluent Tableflow, Microsoft OneLake), `SAPHANA`, `SNOWFLAKEOPENCATALOG`, and `UNITY`; `domain` recipe field now actually emits a `Domains` aspect; stateful incremental ingestion, incremental lineage / properties, profile-skip; synthetic `created = epoch 0` no longer emitted when Dremio doesn't report one; `remove_stale_metadata` and `fail_safe_threshold` exposed.
- **Athena** — S3 Tables (Iceberg) support.
- **dbt** — `skip_missing_upstreams_in_lineage` config; column-level lineage restored for two-tier warehouses (catalog-prefixed SQL + v2 schema fieldPaths); test assertion entities emit an `ownership` aspect when the dbt test node has explicit owner metadata.
- **GCS** — `workload_identity` auth type for GKE Workload Identity; `list_objects_v2` to fix `PaginationError` on Hive-partitioned paths.
- **MSSQL** — consolidated to a single `SqlParsingAggregator`.
- **Teradata** — exponential backoff on transient errors; nullable / autoincrement hydration corrected for `CHAR(N)`-padded values.
- **Redshift** — fixed late-binding view columns silently dropped due to wrong WHERE clause column name.
- **PowerBI** — paginated reports with embedded RDL datasources emit lineage; CTE alias no longer leaks as upstream in native SQL lineage.
- **LookML** — graceful handling of git clone failures and configurable clone timeout.
- **Mode** — `report_pattern` (AllowDenyPattern) config; chart fetch gated on `chart_count` rather than `explorations_count`.
- **SAC** — query Resources OData endpoint directly instead of via `$metadata`.
- **Kafka Connect** — fix duplicated schema segment in sink lineage URNs.
- **Dataplex** — clearer GCP permission error in project-number resolution.
- **Confluence** — page-body HTML converted to Markdown via `markdownify`.

**Ingestion infrastructure:**

- **Per-connector CLI version matrix** with resolution stamp; fall back to default CLI version when the configured version is unset.
- **Patch-based writes for user-editable aspects** — finer-grained partial updates from ingestion.
- **Great Expectations profiler is now optional** — default profiling switches to SQLAlchemy. `acryl-datahub` installs the GE extras only when explicitly requested.
- **`sqlglot[c]` tokenizer restored** on 30.8.0 for performance.
- **Two-tier stored procedure ingestion** — correct URN format and lineage.
- **Skip empty columns for CLL** — avoids spurious column-level lineage entries.

**Executor:**

- **Opt-in ingestion-log garbage collector.** Setting `DATAHUB_EXECUTOR_LOG_GC_ENABLED=true` makes both the remote executor and the coordinator's embedded worker scan `/tmp/datahub/logs/` on an hourly tick and delete per-execution subdirectories older than 14 days. A size cap (default 10 GB) and a 1-hour in-flight grace window apply. Default is `false` for this release; expect the default to flip in a follow-up release.
- **mTLS client authentication** for outbound HTTPS from the executor.
- **Coordinator monitor-request handling** — duplicate monitor requests skipped; polling log chatter trimmed.

#### Breaking Changes

- **Search V2.5 is now enabled by default.** Instances that have not set `SEARCH_VERSION_V2_5_ENABLED` will use V2.5 after upgrading. Existing instances still on legacy V2 may reindex search indices with the V2.5 analyzers during the upgrade. **Migration:** no action required to use the new default. To temporarily roll back to legacy V2, set `SEARCH_VERSION_V2_5_ENABLED=false` for GMS and the system-update job.
- **Dataset semantic search is no longer enabled implicitly.** Enabling semantic search with the default `ELASTICSEARCH_SEMANTIC_SEARCH_ENTITIES=document` no longer also bridges datasets; `dataset` must be listed explicitly (`document,dataset`). **Action:** for instances relying on dataset semantic search, set `ELASTICSEARCH_SEMANTIC_SEARCH_ENTITIES=document,dataset` on GMS and the system-update job.
- **GMS rate limiting renamed.** `rateLimits.defaultRetryAfterSeconds` / `RATE_LIMITS_DEFAULT_RETRY_AFTER` renamed to `minRetryAfterSeconds` / `RATE_LIMITS_MIN_RETRY_AFTER`. The value is now the **minimum** `Retry-After` floor; endpoint (token-bucket) denials may return a longer wait. Added `retryAfterJitterPercent` / `RATE_LIMITS_RETRY_AFTER_JITTER_PERCENT` (default `10`) to spread endpoint retry timing.
- **Airflow plugin: Airflow 2.x dropped.** `acryl-datahub-airflow-plugin` now requires Airflow 3.0+. The plugin always uses `apache-airflow-providers-openlineage` (`>=2.1.0`); drop `openlineage-airflow` from constraints. `[airflow2]` install extra removed; `[airflow3]` retained as a no-op. `taskinstance` URL format and `patch_snowflake_schema` config removed. Pin `acryl-datahub-airflow-plugin <= 1.6.0` for Airflow 2.7–2.10.
- **Prefect plugin: Prefect 3.x required** (`>=3.0.0,<4.0.0`). Entry point group changed from `prefect.block` to `prefect.collections`. Re-register the DataHub block before upgrading.
- **PowerBI Report Server: `chart_pattern` removed** — emits a deprecation warning if set; chart-level filtering is not yet implemented for this connector.
- **Hex: Components ingested as Charts instead of Dashboards.** A Hex Component defines its own visualisation that importing projects cannot override, so it maps to a Chart (analogous to a Looker Look or PowerBI Tile). Component URNs change entity type — saved views, glossary/tag/ownership assignments, and policies that targeted the old Dashboard-typed Component URNs **must be manually reapplied** to the new Chart URNs. Stateful-ingestion stale-removal handles most soft-deletes; component-heavy workspaces may need a one-time bulk cleanup.
- **Document entity: CRUD privileges plus `MANAGE_DOCUMENTS`.** Document CREATE accepts standard entity create privileges (`CREATE_ENTITY` or `EDIT_ENTITY`) or `MANAGE_DOCUMENTS`. UPDATE / DELETE still accept `EDIT_ENTITY` / `DELETE_ENTITY` (or `MANAGE_DOCUMENTS`). Bridge documents inherit VIEW from the source entity; writes do not inherit.
- **Executor coordinator env flags.** `DATAHUB_EXECUTOR_MONITORS_ENABLED` and `DATAHUB_EXECUTOR_TASKS_ENABLED` are now hard opt-outs (skip subsystem wiring and heavy imports), not fetcher-only toggles. Use `DATAHUB_EXECUTOR_INGESTION_PIPELINE_ENABLED` to disable the Kafka / `datahub-actions` pipeline.
- **`getSecretValues` GraphQL query now requires system-level authentication.** The `MANAGE_SECRETS` privilege check remains in place, but `SecretService` now also enforces system-actor auth. Components that fetch secrets at runtime use system credentials in standard deployments and are unaffected. Customers who configured these services with a user-issued PAT must migrate to system credentials before upgrading.
- **Agent env-flag gating now drives `Status.lifecycleStage`.** A SYSTEM agent whose `AI_AGENT_<NAME>_ENABLED` env-var resolves to false is now marked `ARCHIVED` rather than hidden by a read-time filter. Direct URN fetches resolve normally; the existing lifecycle-stage filter keeps ARCHIVED agents out of default search. A startup reconciler restores the previous stage when the flag flips back. Callers that depended on the "null entity" behavior should switch to checking `Status.lifecycleStage = ARCHIVED`.
- **Removed `REQUEST_MINIMAL_SLACK_PERMISSIONS` feature flag.** Replaced by `DATAHUB_SLACK_SERVER_SIDE_HISTORY_ENABLED`, which marks the Slack `:history` scopes as optional in the install screen via `bot_optional`. Admins can deselect them per install rather than the choice being baked in at deploy time.
- **Subscriptions without explicit notification settings now inherit defaults dynamically.** GraphQL callers that omit `notificationConfig` when creating a subscription through `syncSubscription`, or send `notificationConfig` without `notificationSettings`, will use the actor's current notification defaults at delivery time. Callers that intentionally want a no-sink subscription should send `notificationConfig.notificationSettings.sinkTypes: []` explicitly.
- **Action Workflows v2 — field-visibility and validation changes.** For workflows built in the Action Workflows v2 Private Beta, existing definitions are migrated automatically: the older single-field visibility rule is replaced by the standard filter model, and form-field validation is now regex-based (a pattern plus an error message). No action is required. Contact the DataHub Cloud team if you are in the beta and have questions.

#### Deprecations

- **Removed `ENABLE_BEDROCK_OPTIMIZED_LATENCY`.** AWS latency-optimized inference is only available for Claude 3.5 Haiku — not the newer model families (Haiku 4.5, Sonnet 4.x, Opus). The flag was a no-op for those models while inflating every Bedrock cost estimate by 25%. Remove the variable from any deployment config.
- **Dynamic ownership reassignment for proposals is now opt-in.** Proposals continue to work as expected; existing asset owners still receive and can act on proposals. Workflows that depend on ownership reassignment automatically updating who sees proposals must enable the option in Automations.
- **Hex: legacy lineage recipe fields removed** — `lineage_start_time`, `lineage_end_time`, and `datahub_page_size` emit a deprecation warning if set. Lineage now comes directly from the Hex REST API. Remove them from your recipe.

#### Security / Dependencies

- See [updating-datahub.md](../../how/updating-datahub.md) for the full OSS dependency changelog. SaaS-specific notes:
- **`SecretService` caller guard** (see Platform).
- **Removed `REQUEST_MINIMAL_SLACK_PERMISSIONS`** in favor of `DATAHUB_SLACK_SERVER_SIDE_HISTORY_ENABLED` for per-install Slack scope opt-out.
- **Sanitized API error responses** to prevent CWE-200 information leaks.
- **CVE coverage** — `ujson >= 5.12.1` (CVE-2026-44660); `idna >= 3.15` (CVE-2026-45409); plus ongoing dep-bump coverage in line with the OSS changelog.

#### Bug Fixes

- Fixed dataset semantic search bridge backfill robustness during staged rollouts.
- Fixed executor read of prefixed `DATAHUB_EXECUTOR_EMBEDDED_WORKER_ENABLED` env var.
- Fixed `AssertionRunEvent`, `MonitorSuiteInfo`, `SubscriptionInfo`, and `AssertionAssignmentRuleInfo` `schemaVersion` bumps for forward compatibility.
- Fixed Context Hub — prevent duplicate publish context-doc proposals on multiple context-generation runs.
- Fixed CSS selector escape in SchemaTable to prevent `SyntaxError` on special characters.
- Fixed lifecycle-stage migration on documents.
- Fixed semantic-index marker migration to patch existing embedded dataset bridge chunks without requiring a Python re-embed.
- Fixed Display all group relationships on the user profile page.
- Fixed Ask DataHub bold rendering, mode-selector glow, follow-up suggestions.
- Fixed built-in agents from seeing system-agent backend tools.
- Fixed lifecycle-stage issues across glossary terms and documents.
- Fixed missing `applications` field on glossary term and data product GraphQL queries.
- Fixed wrong namespace in `EntityDropdown` (i18n).
- Fixed executor `WorkEventProducer` fail-fast init when a Kafka channel is required.
- Fixed `@KafkaMessagingEnabled` annotation on `KafkaAdminServiceFactory` for pgQueue compatibility.
- Fixed an issue where falsy assertion parameters (ex. row count = 0) caused assertion errors.

## Known Issues

- TBD

#### Environment Variables

- `SECRET_SERVICE_CALLER_GUARD_MODE` (default: `ENFORCE`) — Controls how `SecretService` responds when a non-system actor attempts to decrypt a secret. Values: `ENFORCE` (throw `SecurityException`, recommended for production), `AUDIT` (allow but log a warning — staged rollout), `DISABLED` (no enforcement — break-glass).
- `DATAHUB_EXECUTOR_LOG_GC_ENABLED` (default `false`), `DATAHUB_EXECUTOR_LOG_DIR` (default `/tmp/datahub/logs`), `DATAHUB_EXECUTOR_LOG_GC_INTERVAL_SECONDS` (default `3600`), `DATAHUB_EXECUTOR_LOG_GC_RETENTION_DAYS` (default `14`), `DATAHUB_EXECUTOR_LOG_GC_MAX_DIR_SIZE_MB` (default `10000`; `0` to disable the size cap), `DATAHUB_EXECUTOR_LOG_GC_IN_FLIGHT_GRACE_SECONDS` (default `3600`): Opt-in in-process ingestion-log garbage collector on remote executors and the coordinator's embedded worker.
- `DATAHUB_EXECUTOR_ALLOW_CALL_STATEMENTS` (default `false`): When `true`, custom SQL assertions may use a stored-procedure `CALL` statement in addition to read-only queries. Requires an executor restart.
- `DATAHUB_USE_INFERENCE_V2` (default `false`): Routes smart-assertion training to the V2 pipeline. Requires `DATAHUB_USE_OBSERVE_MODELS=true`.
- `DATAHUB_EXECUTOR_ENABLE_DELTA_BOUNDS` (default `false`): Enables differenced (`boundsValueSpace=DELTA`) prediction bounds in the V1 smart-assertion trainer.
- `ELASTICSEARCH_INDEX_ENTITYMAPPINGLIMITS_<ENTITY>_<LIMIT>`: Per-entity OpenSearch / Elasticsearch mapping ceilings. Use `DEFAULT` as the entity name for a fallback applying to all entity indices.
- `MCP_TELEMETRY_CAPTURE_PAYLOADS`: Enables payload capture for MCP tool invocations surfaced in the MCP Audit tab.
- `RATE_LIMITS_MIN_RETRY_AFTER` (replaces `RATE_LIMITS_DEFAULT_RETRY_AFTER`) and `RATE_LIMITS_RETRY_AFTER_JITTER_PERCENT` (default `10`): GMS rate-limit floor and jitter spread.
- `AUTH_GMS_SESSION_COOKIE_NAME` (default `SESSION`): Name of the Spring Security session cookie set by GMS. Set this if your deployment overrides `spring.session.servlet.cookie.name`.