Connector Improvement: Missing DELETE Events in Oracle Log-Based CDC

Hello,
we are encountering an issue with our Oracle log-based connector:
DELETE operations in the source database are not captured, resulting in stale rows remaining in the destination (Databricks), although they no longer exist in Oracle.

This leads to row-count discrepancies (small in most tables, but consistently reproducible).

Below is a precise description of the issue and evidence collected.

Summary of the Issue

We operate an Oracle SE database and use a Fivetran log-based connector with archived redo logs.

However, in the Fivetran CDC tables, some rows:

still exist with _fivetran_deleted = false
although they have been deleted from Oracle
and also absent in the Teleport snapshot
meaning the DELETE event was never propagated via Fivetran CDC

Evidence

1. CDC shows rows that no longer exist in Oracle

Example: invoice_item_id = 10316280

Present in Fivetran CDC table (_fivetran_deleted = false)
Not present in Oracle (SELECT … WHERE invoice_item_id=10316280 → no rows)
Not present in Teleport snapshot

Conclusion: Insert was captured; Delete was missed.
We validated ~50 IDs — all are deleted in Oracle, confirming a consistent CDC gap.

2. No TRUNCATE or DDL operations

JOP_INVOICE_ITEM has no DDL since 2023-11-24, far before the affected rows (Nov/Dec 2025).
This rules out TRUNCATE or partition maintenance as the cause.

3. No cascading deletes or triggers

No ON DELETE CASCADE constraints
No triggers on the table
All deletes are plain DML DELETEs, which should appear in redo logs

Therefore the missing deletes point to a CDC issue in redo log processing.

__________________

We suspect that since Oracle relies on redo logs (not datafiles) to store committed changes, a missed or unavailable redo log segment during CDC processing could explain why INSERTs were captured but the corresponding DELETEs were never emitted.

Community