Platform, Configuration & Security¶

Two changes here will stop a 2.0 server from starting or behaving as expected with a 1.13 configuration: the LLM/embedding config move, and the new session limits.

LLM and embedding configuration moved¶

Breaking · Affects: every deployment using semantic search, NLQ, or any LLM feature

Provider configuration is no longer nested inside elasticsearch.naturalLanguageSearch. It moves to a new top-level llmConfiguration block, with embeddings as a sub-section that reuses the same provider credentials.

Removed from elasticSearchConfiguration.naturalLanguageSearch¶

embeddingProvider
maxConcurrentRequests
bedrock.*   (awsConfig, modelId, embeddingModelId, embeddingDimension, maxTokens, temperature, timeoutSeconds)
openai.*    (apiKey, endpoint, deploymentName, apiVersion, modelId, embeddingModelId, embeddingDimension, maxTokens, temperature, timeoutSeconds)
google.*    (apiKey, endpoint, modelId, embeddingModelId, embeddingDimension)
djl.*       (embeddingModel)

What remains under naturalLanguageSearch: semanticSearchEnabled, knnNumCandidatesMultiplier (new, default 2) and a new filterExtractor block (cacheMaxSize 1000, cacheExpiryMinutes 5, maxSampleValues).

The 2.0 shape¶

elasticsearch:
  naturalLanguageSearch:
    semanticSearchEnabled: ${SEMANTIC_SEARCH_ENABLED:-false}
    # Embedding provider/model/credentials now live under llmConfiguration.embeddings

llmConfiguration:
  enabled: ${LLM_ENABLED:-false}
  provider: ${LLM_PROVIDER:-noop}   # noop | openai | azureOpenAI | bedrock | google | anthropic
  maxConcurrentRequests: ${LLM_MAX_CONCURRENT_REQUESTS:-5}
  openai:
    apiKey: ${LLM_OPENAI_API_KEY:-""}
    modelId: ${LLM_OPENAI_MODEL_ID:-"gpt-4o-mini"}
    endpoint: ${LLM_OPENAI_ENDPOINT:-""}          # with deploymentName for Azure OpenAI
    deploymentName: ${LLM_OPENAI_DEPLOYMENT:-""}
    apiVersion: ${LLM_OPENAI_API_VERSION:-"2024-02-01"}
    maxTokens: ${LLM_OPENAI_MAX_TOKENS:-4096}
  bedrock:
    awsConfig:
      enabled: ${BEDROCK_AWS_IAM_AUTH_ENABLED:-true}
      region: ${AWS_DEFAULT_REGION:-""}
      accessKeyId: ${AWS_ACCESS_KEY_ID:-""}
      secretAccessKey: ${AWS_SECRET_ACCESS_KEY:-""}
      sessionToken: ${AWS_SESSION_TOKEN:-""}
    modelId: ${LLM_BEDROCK_MODEL_ID:-"eu.anthropic.claude-haiku-4-5-20251001-v1:0"}
    maxTokens: ${LLM_BEDROCK_MAX_TOKENS:-4096}
  google:
    apiKey: ${LLM_GOOGLE_API_KEY:-""}
    modelId: ${LLM_GOOGLE_MODEL_ID:-"gemini-2.5-flash"}
  anthropic:
    apiKey: ${LLM_ANTHROPIC_API_KEY:-""}
    modelId: ${LLM_ANTHROPIC_MODEL_ID:-"claude-3-5-sonnet-20240620"}
    baseUrl: ${LLM_ANTHROPIC_BASE_URL:-"https://api.anthropic.com"}
  embeddings:
    provider: ${EMBEDDING_PROVIDER:-bedrock}   # bedrock | openai | google | djl
    maxConcurrentRequests: ${MAX_CONCURRENT_EMBEDDING_REQUESTS:-10}
    bedrock:
      embeddingModelId: ${AWS_BEDROCK_EMBED_MODEL_ID:-"amazon.titan-embed-text-v2:0"}
      embeddingDimension: ${AWS_BEDROCK_EMBEDDING_DIMENSION:-512}
    openai:
      embeddingModelId: ${OPENAI_EMBEDDING_MODEL_ID:-"text-embedding-3-small"}
      embeddingDimension: ${OPENAI_EMBEDDING_DIMENSION:-1536}
    google:
      embeddingModelId: ${GOOGLE_EMBEDDING_MODEL_ID:-"gemini-embedding-001"}
      embeddingDimension: ${GOOGLE_EMBEDDING_DIMENSION:-768}
    djl:
      embeddingModel: ${DJL_EMBEDDING_MODEL:-"ai.djl.huggingface.pytorch/sentence-transformers/all-MiniLM-L6-v2"}

The embedding provider may differ from the chat provider; embeddings reuse the credentials from the provider blocks above (bedrock.awsConfig, openai.apiKey/endpoint, google.apiKey).

Environment variable renames¶

AWS_BEDROCK_EMBED_MODEL_ID, AWS_BEDROCK_EMBEDDING_DIMENSION, OPENAI_EMBEDDING_MODEL_ID, OPENAI_EMBEDDING_DIMENSION, GOOGLE_EMBEDDING_MODEL_ID, GOOGLE_EMBEDDING_DIMENSION, DJL_EMBEDDING_MODEL, EMBEDDING_PROVIDER and MAX_CONCURRENT_EMBEDDING_REQUESTS keep their names but move under llmConfiguration.embeddings.

Session management¶

Sessions are database-backed¶

Behavioural · Affects: multi-pod deployments

2.0 adds a user_session table so sessions survive pod restarts and are shared across pods. Previously each pod held its own in-memory session state, which caused spurious logouts behind a load balancer without sticky sessions.

Columns: id, userId, status, expiresAt, idleExpiresAt, updatedAt, sessionType, provider, version, lastAccessedAt, refreshLeaseUntil, plus the session JSON.

Concurrent sessions are capped per user¶

Behavioural — users will be logged out

authenticationConfiguration:
  sessionExpiry: ${AUTHENTICATION_SESSION_EXPIRY:-"604800"}   # 7 days, all auth providers
  maxActiveSessionsPerUser: ${AUTHENTICATION_MAX_ACTIVE_SESSIONS_PER_USER:-5}

Maximum number of active authenticated sessions allowed per user. When the limit is exceeded, the least recently used active sessions are revoked. If unset, OpenMetadata uses the default of 5.

sessionExpiry now applies to all auth providers (minimum 3600 s). oidcConfiguration.sessionExpiry becomes a deprecated fallback.

Additional trusted redirect URIs¶

additionalTrustedRedirectUris: ${AUTHENTICATION_ADDITIONAL_TRUSTED_REDIRECT_URIS:-[]}

Allows redirect URIs beyond callbackUrl and the server's own callbacks — each entry must match the requested URI exactly (scheme, host, port, path, query). Intended for browser-extension logins, e.g. https://<extension-id>.chromiumapp.org/auth0.

Related SSO hardening in 2.0: the server /auth/callback is trusted in the SAML redirect allowlist, SAML pending-session ids are carried in RelayState, and the OIDC login-loop is hardened with an interactive fallback on login_required that preserves pending logins.

Admin test-login¶

POST /v1/system/security/test-login/validate-token

Validates a browser-obtained OIDC id_token against a candidate (unsaved) security configuration, so an admin can confirm a real login resolves the expected identity before saving. Schemas: system/testLoginTokenRequest.json, system/testLoginResult.json.

LDAP¶

recursiveGroupMembership (default false) enables transitive group resolution for Active Directory nested groups via LDAP_MATCHING_RULE_IN_CHAIN.

Database connection timeouts changed¶

Behavioural — long-running queries will now be cut off

Server & logging configuration¶

HTTP/2 is available (opt-in)¶

server:
  applicationConnectors:
    - type: ${SERVER_PROTOCOL:-http}   # http (default) | h2c (cleartext HTTP/2) | h2 (TLS)

dropwizard-http2 is on the classpath, so h2c and h2 connector types work out of the box. Both are backwards compatible — HTTP/1.1 clients keep working on the same port. Worth enabling when browsers hit Jetty directly (Docker Compose, on-prem single node); not worth it behind an HTTP/2-terminating load balancer.

Response compression enabled¶

server.gzip.enabled is true and the comment changes from "Response compression disabled for maximum throughput" to a description of Jetty's GzipHandler. Responses above ~256 bytes are now gzipped. Clients that mishandle Content-Encoding: gzip — rare, but possible in hand-rolled HTTP code — will need Accept-Encoding: identity.

Logging hardening¶

Security-relevant

logging:
  loggers:
    org.eclipse.jetty:
      level: ${JETTY_LOG_LEVEL:-INFO}
    org.openmetadata.service.audit.AuditLogRepository:
      level: INFO
  appenders:
    - type: console
      filterFactories:
        - type: audit-exclude-filter-factory

LOG_LEVEL sets the root logger, and at DEBUG Jetty's HttpParser prints every request header verbatim — including Authorization: Bearer <jwt> and Cookie: OM_SESSION=<id>. Since DEBUG is exactly what support asks customers to enable, and those logs get attached to tickets, the Jetty logger is now pinned independently via JETTY_LOG_LEVEL.

Audit entries are logged at INFO with an AUDIT marker and routed to logs/audit.log; the level is pinned so audit capture does not depend on the global LOG_LEVEL, and an audit-exclude-filter-factory keeps them out of the console appender.

Object storage configuration expanded¶

· Affects: deployments using file attachments

  objectStorage:
-   enabled: false
-   provider: NOOP
-   maxFileSize: 5242880
+   enabled: ${ASSET_UPLOADER_ENABLE:-false}
+   provider: ${ASSET_UPLOADER_PROVIDER:-s3}   # s3 | azure | inmemory
+   maxFileSize: ${ASSET_UPLOADER_MAX_FILE_SIZE:-5242880}
+   s3: { endpoint, bucketName, region, accessKey, secretKey, useIamRole, iamRoleArn,
+         prefixPath, sseAlgorithm, kmsKeyId }
+   azure: { containerName, connectionString, useManagedIdentity, clientId, tenantId,
+            clientSecret, blobEndpoint, prefixPath }

The default provider changes from NOOP to s3 — but enabled still defaults to false, so nothing activates until you turn it on. This backs the new /v1/attachments API (uploaded assets for Context Center files, Knowledge Center pages and cover images). For MinIO, use provider s3 and point s3.endpoint at the MinIO server.

Query timeouts and workflow polling¶

See Data Governance → Workflows fire far sooner for the Flowable interval changes, which raise steady-state database polling.

Ingestion pipeline client¶

pipelineServiceClientConfiguration gains queuedStatusTimeoutSeconds (default 3600) — how long a queued status stays visible before being treated as stale and hidden. Covers runs an orchestrator accepted but never started.

The 2.0.0 database migration¶

Plan a maintenance window

New tables¶

Altered tables¶

• background_jobs — adds progress, total, result, error, message, cancelRequested, completedAt
• tag_usage — adds metadata JSON
• audit_log_event — adds search_text LONGTEXT
• thread_entity → renamed to thread_entity_legacy

Data migrations¶

Index additions¶

Many, including (deleted, name) and (deleted, serviceType) composites on all 13 service tables (for /v1/services/overview), name indexes on the new entity tables (so the distributed reindex cursor runs index-only instead of a filesort that can exhaust sort memory on large tables), and idx_wf_instance_state_execution_id on workflow_instance_state_time_series.

Dependency and CVE updates¶

— not breaking, but relevant to hardened deployments.

Backend: Jetty 12.1.7 → 12.1.10 (with jetty-bom imported so transitive modules follow), Netty 4.1.137.Final, BouncyCastle 1.85, jackson-databind 2.18.8, tools.jackson BOM 3.1.5, log4j 2.25.5, libthrift/thrift 0.24.0, reactor-netty-http 1.2.18, tomcat-jdbc/juli 11.0.11, httpcore5 5.4.3, Redshift JDBC 2.1.0.30 → 2.2.2, Kubernetes client-java 25.0.1, Apache Airflow 3.3.0 on Python 3.12.

Frontend: ws 8.21.0, handlebars 4.5.2, js-yaml 5.2.2, fast-uri 3.1.5, nanoid 3.3.17, brace-expansion 1.1.18 / 5.0.9.

Ingestion images: apt package trimming for gnutls, libcap, openssh, rsync, util-linux and the Go CVE surface; mlflow-skinny and pyarrow bumped.

Other security fixes worth noting operationally:

• Test-connection workflow triggers are authorized.
• CSRF failures fail-secure and retry on the next request instead of permanently breaking.
• POST .../testDestination redacts destination config from the response.
• SCIM displayName sync is fixed (a bot-PUT guard was reverting it).