Oracle CDC Benchmark Results

Benchmark suite: internal/impl/oracledb/bench/

2026-03 — Initial Benchmark

PR: https://github.com/redpanda-data/connect/pull/4082

Environment:

Oracle Database Express Edition (XE) 21c running in Docker
Container image: container-registry.oracle.com/database/express:latest

Dataset: 18,000,000 rows, ~3GB

sql

SELECT segment_name, bytes / 1024 / 1024 AS mb FROM dba_segments WHERE owner = 'TESTDB';

SEGMENT_NAME         MB
_______________ _______
USERS              2950
SYS_C008322         288

Configuration:

snapshot_max_batch_size: 160000
logminer.scn_window_size: 190000
batching.count: 140000

Snapshot Results

Metric	Value
Messages/sec	~140,000

Snapshot mode benefits from greater concurrency and behaves similarly across SQL-based connectors.

LogMiner Streaming Results

Metric	Value
Total runtime	~12 minutes
Average throughput	~50,000 msg/sec (total average: 25,000 msg/sec)
Peak throughput	~70,000-90,000 msg/sec

Observations:

LogMiner was never designed for CDC and has inherent performance limitations:
- Requires buffering transactions until COMMIT or ROLLBACK before flushing
- Large gaps between redo log SCNs before reaching relevant data
- Single-threaded by nature
Removing the buffering layer improved things slightly but not significantly, confirming the bottleneck is LogMiner fetch speed
Quick comparison with Debezium's Oracle CDC (also LogMiner-based) showed similar throughput, confirming this is a protocol limitation, not a Connect limitation