Redpanda Connect v4.101.0 - CDC Isolation and Row Integrity

Redpanda Connect v4.101.0 was published on July 17, 2026. The main operational change is a Snowflake streaming check intended to prevent silent data corruption when Parquet metadata disagrees with the rows actually serialized.

The full release notes and downloads are on the GitHub release page.

The aws_dynamodb_cdc input now accepts an optional checkpoint_namespace field. It allows multiple independent pipelines to use one DynamoDB checkpoint table without overwriting each other’s state. This is a small configuration addition with a direct effect on recovery correctness.

Previously, sharing the checkpoint table could create state collisions between pipelines. That is dangerous during restart or replay because a pipeline relies on its stored position to decide where processing resumes. The namespace gives operators a way to separate those positions while retaining one physical table.

The field is optional, so existing configurations do not need an immediate edit. Operators consolidating checkpoint storage should still assign a distinct namespace to every independent pipeline and keep that value stable across deployments. Changing it effectively points the pipeline at a different checkpoint identity. The release notes do not describe automatic namespace assignment, validation across deployments, or migration of existing checkpoint records, so those details remain operator responsibilities. The implementation is tracked in pull request 4602.

The oracledb_cdc input adds snapshot_filters for SQL SELECT overrides on individual tables during initial snapshot capture. This gives data teams control over which rows enter the baseline without requiring a separate extraction job or a broad filter applied to every table.

The scope matters. The release note ties this setting to the initial snapshot, not the continuing change stream. A filtered baseline therefore needs to remain consistent with the events that arrive after snapshot completion. If later changes can introduce rows excluded by the initial query, downstream consumers must be able to handle records that have no prior snapshot image.

Per table SQL also moves some correctness responsibility into pipeline configuration. Query review, Oracle permissions, and repeatable deployment values deserve the same controls as schema mappings. The new option is documented by the change in pull request 4606.

The snowflake_streaming output now verifies that the Parquet row count matches the number of rows actually serialized before upload to Snowflake. The project describes the fixed failure mode as potential silent data corruption. That makes this the release item to prioritize for pipelines where completeness matters more than merely observing a successful upload.

The check sits at the boundary between in memory records and the generated Parquet payload. A mismatch at that point means metadata and physical content disagree before Snowflake receives the object. Detecting the discrepancy before upload gives the connector a chance to avoid treating an incomplete payload as valid.

The notes do not specify the conditions that produced a mismatch, nor do they document new configuration or migration steps. Operators should treat this as a correctness fix rather than a tuning feature. After upgrading, monitor existing error and retry signals for Snowflake outputs, especially on workloads with varied schemas or batch sizes. The fix is in pull request 4605.