Skip to main content

CDC-7001 Orphan release violation

Code: CDC-7001 (ORPHAN_RELEASE_VIOLATION). Not retried: the task stops at once.

What the connector observed​

LogMiner returned a COMMIT for a transaction whose buffered changes the connector had already let go of, in one of two ways:

  • Orphan detection released it as rolled back (cdc.transaction.orphan.action=release, the default), because the transaction was absent from GV$TRANSACTION in two checks, the connector had mined past the first absence, and the session named by its START row no longer existed.
  • The long-transaction policy discarded it (cdc.transaction.max.age.action=discard) because it was open longer than cdc.transaction.max.age.ms.

The ids of released and discarded transactions travel in the connector's offsets (released_xids, the most recent 256), so this is detected after a restart too. The message names the transaction id and the COMMIT SCN.

Why it stopped rather than continued​

When the transaction was released or discarded, its earlier changes were dropped. Publishing the COMMIT now would publish only the changes mined after that point: a partial transaction, which is worse than none. The connector stops instead.

Confirm the cause​

The ops topic holds the decision: a transaction-orphan-released event (with first_scn, last_scn, absent_at_scn and reason) or a transaction-discarded event (with first_scn, last_scn and age_ms) for the same transaction id. A discarded transaction also has a record on the DLQ topic. Compare the event's SCNs with the COMMIT SCN in the message.

A released transaction that later commits means the database still had it while the probe did not see it, for example because the transaction ran on an instance or in a container the connector user cannot see in GV$TRANSACTION. Check the user's CONTAINER_DATA with oracle-cdc-doctor check (rule DOC-4).

Recover​

To publish the transaction whole, mine it again from its first change:

  1. Stop the connector and read its offset (moving the offset by hand).
  2. Set resume_scn to the first_scn from the ops event (or lower), remove resume_rs_id and resume_ssn, and remove the transaction's id from released_xids. Keep the last_commit_* fields, so transactions already delivered are skipped.
  3. Write the offset back and resume. The redo from first_scn must still be available.

oracle-cdc-admin offsets set does these steps in one command, checks that the redo is present and records the change on the ops topic:

java -jar oracle-cdc-doctor-cli.jar admin offsets set --connect-url "$CONNECT" --name "$NAME" \
    --scn <first_scn> --forget-released <id from released_xids> \
    --reason "re-mine the transaction CDC-7001 named" --resume

If that redo is gone, the transaction cannot be published whole. Move the offset past its COMMIT SCN as a deliberate skip, remove the id from released_xids, and reload the tables the transaction touched with a snapshot signal.

To make a repeat less likely, raise cdc.transaction.orphan.check.interval.ms or set cdc.transaction.orphan.action=fail, so that the next suspected orphan stops the task (CDC-7002) for a decision instead of being released; and raise or unset cdc.transaction.max.age.ms if the discard policy caught a legitimate transaction.