Skip to main content

​
Applications & Automation

Four installed applications have their stored configuration rewritten by the 2.0.0 migration: Search Indexing, RDF Index, Data Insights and MCP. If you manage application configuration as code, read this before you replay it. The Documentation, Tier, and Quality agent apps are deleted by the same migration — that’s covered separately under AI Agents & Automations.

​
Search Indexing app configuration fields are removed

Breaking. Affects scripted reindex triggers and infrastructure-as-code.
In 2.0, reindex always recreates the index and distributed indexing is always used — there is no incremental mode. New supporting infrastructure: search_index_job for distributed reindex job tracking, and search_index_retry_queue for failed search-index writes with claim leasing.
Any automation triggering a reindex with {"recreateIndex": false} — an “incremental reindex” job, for example — must drop the field. New diagnostics ship alongside: reindex-drift and index-health checks in /v1/system/validate, plus a cluster-fitness diagnostic.

​
RDF Index app defaults change, and your settings get overwritten

Behavioural. The migration overwrites your settings, not just the schema defaults. Incremental RDF indexing produced unbounded triple growth because relationship-removal paths weren’t fully reconciled. A full per-run clear always converges to the relational state, and the weekly cadence keeps per-run cost from saturating the triple store. The entities reset is a safety measure: with a full rebuild issuing a clear before indexing, a narrowed entity subset would wipe triples for entity types still present in the database.
Your entities subset is not preserved. Reapply it after the migration if you need partial RDF indexing — understanding that the following run will still clear first. inferenceEnabled also flips from true to false; set RDF_INFERENCE_ENABLED=true to restore 1.13 behaviour.

​
Data Insights and MCP application configuration are stripped

Breaking, for two unrelated reasons landing in the same migration pass. A stored DataInsightsApplication configuration containing moduleConfiguration.dataQuality fails to deserialize at startup — see Data Quality for the full removal. Separately, MCP configuration now lives solely in the mcpConfiguration setting; the application-level copy (which no code read) is dropped by the migration, and the now-empty configure step is hidden in the UI. Configure MCP through Settings > MCP, not through the application’s own configuration. Two more fields are stripped from stored application rows for security reasons (behavioural): openMetadataServerConnection and privateConfiguration were being persisted into application rows and version history despite being runtime-only. 2.0 also stops exposing the application bot JWT and privateConfiguration secrets in API responses — if your tooling read the bot token out of /v1/apps/..., obtain it through the bot API instead.
The appConfig union also drops the three Collate AI agent configurations and gains a permissive terminal branch — branch indexes shift, so generated unions that key on position must be regenerated rather than hand-patched.

​
MCP tool contracts change for 15 tools, and 7 tools are added

Breaking for MCP clients that parse responses. No tools are removed, but the contract changes in ways that will silently misbehave if you don’t update the client: The changed tools are create_classification, create_data_product, create_domain, create_glossary, create_glossary_term, create_lineage, create_metric, create_tag, create_test_case, get_entity_lineage, get_test_definitions, patch_entity, root_cause_analysis, search_metadata and semantic_search. The seven new tools are all Context Center and AI-context tools: create_context_memory, find_context, get_asset_context, get_company_context, get_knowledge_content, get_persona_context and search_company_context.
Re-run your MCP client against 2.0 before rolling out. Anything that counted results with a size: 0 aggregation query, or paged with a bespoke offset, must change.

​
CSV import and export become tracked background jobs

Behavioural. Affects CSV automation and the UI. In 1.13, async CSV progress was delivered only over WebSocket — a client without a live socket had no way to observe the job. In 2.0, background_jobs gains progress, total, result, error, cancelRequested and completedAt; jobs are pollable at GET /v1/csvAsyncJobs/{jobId} and downloadable from GET /v1/csvAsyncJobs/{jobId}/result, and the UI surfaces them in a Background jobs tray. Async entity restore is also available via PUT /v1/{entityType}/restore?async=true.

​
Concurrency limits protect the connection pool during background fan-out

Additive. A new openmetadata.yaml block caps how much background work can run at once:
This prevents background fan-out (reindex, Data Insights runs) from exhausting the request connection pool. Tune the values down if you see connection-pool starvation during those runs.

​
Search and vector infrastructure gain new capabilities

Additive. Native Elasticsearch vector search is available alongside the existing OpenSearch path, and Google Gemini joins as an embedding provider. Multi-chunk hybrid ranking is backed by denormalised chunk documents with a staged, generation-based index recreate, inspectable at GET /v1/search/vector/fingerprint. The candidate multiplier for vector search is tunable via knnNumCandidatesMultiplier (num_candidates = max(k × multiplier, 100), default 2), and NLQ’s filterExtractor block gains its own cache size, TTL, and prompt sample limits.

​
A new cache layer ships with admin controls

Additive. New admin endpoints — GET /v1/system/cache/keys, POST /v1/system/cache/invalidate and POST /v1/system/cache/invalidate/entity — manage a cache layer that’s a no-op by default. Set CACHE_PROVIDER=redis to enable the Redis L2 cache; it falls back to the no-op provider if Redis is unreachable at startup.

​
Full detail

For the full narrative walkthrough of Platform Applications in Collate 2.0 — including screenshots and step-by-step context — see Release 2.0: Platform Applications.