Redpanda Connect v4.110.0 - Snowflake Offsets and Oracle CDC

Redpanda Connect v4.110.0 was published on 17 September 2026 at 22:33 UTC. The fix operators will notice is offset token comparison in snowflake_streaming. Unpadded numeric tokens were ordered incorrectly once the digit count increased, which can write duplicate rows or skip deliveries.

The full release notes and downloads are on the GitHub release page. Install steps are still in the getting started guide. The accumulated history is in the change log.

snowflake_streaming tracks load progress with an offset token. Pull request 4816, from Jeffail, corrects comparison of unpadded numeric tokens.

A numeric token stored without padding grows an extra digit when it crosses a power of ten. 999 and 1000 are the usual pair. Compared as text, 999 sorts after 1000, because the character 9 is greater than 1. Compared as integers, 999 is the smaller value. A streaming output that treats the text order as “newer” will pick the wrong token once the digit count changes.

The release note names both outcomes of that bad order: duplicate rows, and missed deliveries. A reversed comparison either replays a range already committed or jumps past a token that was never flushed. The published fix is numeric order for unpadded tokens.

Channels that already crossed a digit boundary on an older build can still hold bad rows. This tag stops the comparison from repeating the mistake. Rows already landed in Snowflake stay there. Keep a dedupe key on the target table when the stream can replay.

The Snowflake integration moves its driver from v1.19.0 to v2.2.0 in pull request 4822, from squiidz. The note gives two reasons. The bump fixes CVE-2026-85525. It also removes the deprecated RaisePutGetError option.

v1 to v2 is a major driver bump. The release note does not itemize the rest of the driver diff. What this tag publishes is the CVE fix and the deleted option. Treat any other behavior change as unlisted until you measure it on your own queries.

CVE-2026-85525 is the only security identifier in the note. Severity, the affected call path, and the upstream commit are not copied into the release text. Look that identifier up in your advisory feed when Snowflake is in the exposure set. A process that stays on driver v1.19.0 keeps the issue.

This driver drops RaisePutGetError. The name refers to Snowflake PUT and GET error handling. The note only says the option was deprecated and is now gone. There is no replacement key in the release text.

oracledb_cdc picks up two changes, both from josephwoodward. Test them as one connector upgrade.

Pull request 4674 adds the Oracle username to event metadata. Traceability is the stated reason. A pipeline fed by more than one database account can separate those accounts on the event. Audit jobs can read the connecting user from metadata. The release note does not publish the metadata key name. Sample one event on a non production stream and bind mappings to the key you actually see.

Pull request 4821 switches the same connector to prepared statements. The note cites performance and resource management. Prepared statements skip a fresh parse for SQL the connector sends repeatedly, and they give Oracle a stable cursor to reuse. Each prepared handle is session state. It occupies memory on the database session until the connector closes it.

Oracle session views that matched literal SQL will show the prepared text after this tag. Update DBA scripts that grep those views for the previous string.

The Iceberg line is a test fixture change. Pull request 4819, from Leward, replaces MinIO with a silo fork in tests. MinIO images were removed from Docker Hub, so a CI job that pulled the old image could not start the object store used by the Iceberg tests.

The Iceberg connector runtime is outside this pull request. Production jobs still talk to the object store named in your config. The silo fork exists so the project test suite can start. Image names in your own pipelines are unchanged by v4.110.0.

Delete RaisePutGetError from Snowflake configuration before you deploy v4.110.0. Driver v2.2.0 no longer includes that deprecated option. The release note does not describe a replacement setting. Force one failing PUT or GET in a test pipeline and confirm the error still shows up in logs.

snowflake_streaming checkpoints that use unpadded numeric tokens now follow numeric order. Watch the first checkpoint after upgrade when a token sits near a change in digit count. Compare source counts with the Snowflake table for that range. The release includes no repair tool for rows already written.

oracledb_cdc adds username metadata and sends prepared statements. No migration command is listed. Refresh mappings that assumed a fixed metadata map. Refresh any Oracle session query that matched the previous SQL text.