VERA 4.6 — 2026-08-18

Compute Engine: VERA 4.6

Built-on: Apache Flink® 1.20

Overview

VERA 4.6 introduces multimodal AI processing for real-time inference on images and PDFs, a new VARIANT type for semi-structured data, and CDC YAML data ingestion reaching general availability. This release also bundles two Ververica Cloud platform features: Single Sign-On with Organizations for centralized team access management, and Audit Logs with encryption-key management for security and compliance. It includes lakehouse improvements and community bug fixes from Apache Flink® 1.20.2/1.20.3.

New Features

Multimodal AI Function Support

VERA 4.6 adds built-in functions for image and PDF preprocessing ahead of Vision Language Model inference: FETCH_CONTENT, MIME_TYPE, IMAGE_COMPRESS, IMAGE_LAPLACE_VAR, and PDF_TO_IMAGES. CREATE MODEL adds a content_type parameter — image_url is supported in this release, and input_audio and document are recognized but not yet usable. ml_predict, ai_extract, and ai_classify now accept image inputs through a FETCH_CONTENT → TO_BASE64 pipeline.

Benefits:

• Runs real-time inference on images and PDFs with Vision Language Models such as Qwen-VL, without leaving SQL.
• Covers the full pipeline: fetch, detect type, compress, and check sharpness before inference.

Learn More: Multimodal Functions

VARIANT Type: Field Access and Paimon Sink Support

The VARIANT type gains field access syntax (variant.field, variant['key']) and Paimon Sink write support, alongside a new parse_json function that converts a JSON string to VARIANT, and an MD5 function for data ingestion transforms.

Benefits:

• Reads and writes semi-structured data without converting it to a fixed schema first.
• Extends VARIANT support through to Paimon storage.

Learn More: Data Type Conversion

CDC YAML Data Ingestion: General Availability

Data ingestion through CDC YAML exits public beta and is now generally available for the Paimon, StarRocks, MySQL, and Kafka connectors, with MongoDB, Postgres, and Fluss entering public preview.

• Merge multiple columns from an upstream JSON source into a single target column.
• Append-only writes to partitioned Paimon tables without a primary key.
• Clear the inferred PrimaryKey or PartitionKey to convert a table to one without a primary key.
• Complex table-name routing using regular expressions.
• VARIANT full-link support: access, convert, and write VARIANT-type fields through CDC YAML.
• Kafka source field-routing (splitting one message into records for different target tables) and custom partitioners.
• Fluss sink support for database CDC with schema evolution; independently configurable Paimon sink commit concurrency; built-in Iceberg sink catalog references.

Benefits:

• Moves a core data ingestion path from beta to a generally available, supported feature.
• Handles more real-world CDC schema shapes directly in YAML, without custom transform code.

Learn More: Data Ingestion (CDC YAML)

Organizations & Access Control

Single Sign-On and Organizations

Ververica Cloud organizations can now centralize access through Single Sign-On, using either OpenID Connect or SAML 2.0. Converting an individual account into an organization brings Organization Teams for high-level administration, custom Teams mapped to identity-provider groups for workspace access, and guest invitations for external collaborators.

Benefits:

• Manages who can sign in and what they can access from your own identity provider, instead of per-user Ververica credentials.
• Separates organization-wide administration (billing, users, workspaces) from day-to-day workspace access.

Learn More: Single Sign-On Overview

Audit Logs and Encryption Keys

Audit Logs capture platform events for security and compliance, in a pull-based model you can retrieve, query, and retain for as long as you need. Log content is protected with hybrid RSA-4096/AES-GCM encryption, and you manage the encryption keys directly.

Benefits:

• Provides traceability of platform activity to support governance and compliance requirements.
• Keeps audit data encrypted end-to-end, with keys you control.

Learn More: Audit Logs

Connectors

• Kafka sink supports the Debezium JSON three-part table ID (Database.Schema.Table).
• Fixed a Kafka connector issue where a stateful restart after a topic switch could consume both the old and new topics.
• Improved MySQL CDC GTID expiration error messages, and the consumer Server ID now appears in logs.

Lakehouse Integration

• Iceberg sink reports a new numRecordsOutOfSinkPerSecond metric, supports configurable Hadoop parameters, and can be used as a data ingestion target for DLF Iceberg.
• Paimon sink supports the Lance file format.

Observability

• New metrics: local disk usage, Gemini native memory usage/limit, and Auto-pilot operator concurrency limits.
• Reduced log noise by suppressing unnecessary WARN-level messages.

Bug Fixes

The following issues are resolved:

• Merged community fixes from Apache Flink® 1.20.2/1.20.3.
• Fixed Kafka connector data loss when transactions were enabled for Kafka-to-OSS writes.
• Fixed Canal Protobuf format inconsistencies (timestamp format, tinyint type) in CDC YAML.
• Fixed MySQL CDC source reuse showing one fewer table during debugging.
• Fixed the YAML ODPS sink causing a Metaspace out-of-memory error from frequent commits.
• Improved the error message for Temporal Join syntax.