This page is for data engineers, DBAs and architects who need changes to flow out of an operational database continuously rather than in a nightly batch. It covers enterprise log-based replication suites such as Qlik Replicate, Oracle GoldenGate, IBM Data Replication, Fivetran HVR and Tungsten Replicator; streaming integration platforms such as Striim, IBM StreamSets and Estuary; the open source Debezium; cloud services such as AWS DMS and the Confluent CDC source connectors; and bidirectional synchronisation tools such as SymmetricDS.
Ownership and lifecycle have shifted in this market, and the listings reflect it. SQL Data Sync for Azure SQL Database retires on 30 September 2027 and no longer accepts new sync groups in subscriptions that never used it. Arcion was acquired by Databricks in 2023 and is kept as a historical profile, because its technology now sits inside Databricks Lakeflow Connect. IBM now owns both StreamSets (since July 2024) and Confluent (since March 2026), alongside its own IBM Data Replication.
What this software does
Log-based capture is the mechanism most products here are built on. The tool reads the database's own change record: the Oracle redo log, the SQL Server transaction log or its CDC tables, the MySQL binary log, PostgreSQL logical decoding through a replication slot, or MongoDB change streams. It turns each committed insert, update and delete into a row-level change event, in commit order, without querying the source tables. Trigger-based capture (used by SymmetricDS Community and SQL Data Sync) instead adds triggers that copy each change into side tables, which works where logs are not accessible but adds write overhead on the source. Polling a timestamp column is the simplest approach but misses hard deletes and intermediate states, which is why buyers usually rule it out for replication.
A typical pipeline runs an initial load (a snapshot of existing rows), then switches to streaming changes from a recorded log position. The tool stores offsets or checkpoints so it can resume after a restart without gaps, and it has to decide what to do with schema changes (DDL) on the source. Delivery either applies changes directly to a target database or warehouse, or publishes them to a message bus such as Apache Kafka, where other consumers pick them up. Most CDC delivery is at-least-once, so targets and consumers should be idempotent.
Most CDC is one-way. Bidirectional or multi-primary replication, where two databases both accept writes, also needs conflict detection and resolution: Oracle GoldenGate documents automatic conflict resolution for Oracle-to-Oracle replication, SymmetricDS has configurable conflict rules, and SQL Data Sync offers a simple hub-wins or member-wins policy. CDC tools differ from batch ETL, which reloads or queries data on a schedule, and from a database's built-in replication (such as read replicas), which usually only copies between servers of the same engine.
Types of product in this category
- Enterprise replication: Commercial log-based replication engines for heterogeneous estates, including mainframe and SAP sources in some cases (Qlik Replicate, Oracle GoldenGate, IBM Data Replication, Fivetran HVR, Tungsten Replicator), usually sold by quote and often self-hosted.
- Real-time integration platform: Platforms that combine CDC with streaming pipelines, in-flight transformation and many non-database connectors (Striim, IBM StreamSets, Estuary; Arcion historically).
- Data integration platform: General ELT platforms with a CDC mode for database sources, such as Airbyte, where CDC runs as scheduled syncs rather than a continuous stream.
- Open source CDC: Free, self-operated capture software such as Debezium, which streams log-based change events through Kafka Connect, Debezium Server or an embedded Java engine.
- Cloud service: Managed capture services run by a cloud or streaming vendor, such as AWS DMS (full load plus CDC between endpoints) and the Confluent CDC source connectors, which deliver changes into Kafka.
- Database synchronization: Tools that keep two or more databases in sync in both directions with conflict rules, such as SymmetricDS and the retiring Azure SQL Data Sync, typically using triggers.
Typical use cases
Zero-downtime database migration
Load the target once, keep it current with CDC while applications still write to the source, then cut over in minutes. AWS DMS, Oracle GoldenGate and Qlik Replicate all document full load followed by change processing for this pattern.
Real-time feeds into a warehouse or lakehouse
Stream operational changes into Snowflake, BigQuery, Databricks or Fabric so dashboards and models work on minutes-old data instead of last night's extract. Check that the tool merges updates and deletes correctly, not just appends.
Event streaming and microservices
Publish row changes to Kafka topics (Debezium, Confluent CDC connectors) so other services react to orders, payments or profile updates without polling the database or dual writes from application code.
Keeping caches and search indexes current
Use change events to update or invalidate a Redis cache or an Elasticsearch or OpenSearch index when the source row changes, instead of rebuilding them on a schedule.
Mainframe and legacy offload
Replicate Db2 for z/OS, IMS or VSAM data to modern platforms for reporting and new applications. Only some products capture these sources natively, notably IBM Data Replication, Qlik Replicate and Oracle GoldenGate.
Branch, edge and bidirectional sync
Keep store, branch or field databases in sync with a central system over unreliable links, with conflict rules for rows changed in two places. SymmetricDS is designed for this; most one-way CDC tools are not.
How to choose
- Source and target coverage, by version. Check the exact source and target versions you run, not just the engine names. Lists differ widely: Qlik Replicate and IBM cover mainframe sources that Striim does not, Tungsten Replicator only captures from MySQL-family servers, and the Confluent connectors deliver only to Kafka. Endpoint support also ages out; Qlik Replicate retired MySQL 8.0 and Db2 LUW 11.5 sources in May 2026.
- Capture method and source impact. Log-based capture needs specific settings and privileges (supplemental logging on Oracle, row-based binlog on MySQL, a logical replication slot on PostgreSQL), and a stalled PostgreSQL slot keeps WAL on disk until the consumer catches up. Trigger-based capture is easier to enable but adds work to every write. Ask the vendor what it reads and what it installs on the source.
- Initial load and resynchronisation. Look for an initial load that runs without locking the source for long, and for a way to re-snapshot one table without restarting the pipeline (Debezium supports signal-driven incremental snapshots). Several profiles here do not document this clearly, so confirm it in a trial.
- Schema change handling. Decide what should happen when someone adds or drops a column. Debezium records schema history; Striim lets you choose per consumer; Tungsten Replicator does not carry DDL to heterogeneous targets; SQL Data Sync does not replicate schema changes at all. Unhandled DDL is a common cause of broken pipelines.
- Ordering, consistency and delivery guarantees. Check whether changes are applied in transaction order with transaction boundaries kept, and whether delivery is at-least-once (most tools) so the target must tolerate duplicates. For Kafka-based designs, ask how offsets are stored and what happens on connector restart.
- Topology and conflict handling. One-way fan-out, fan-in, bidirectional and active-active need different features. Bidirectional replication requires conflict detection and a resolution policy; Debezium, Confluent and StreamSets are one-way only, while GoldenGate, SymmetricDS and Tungsten document multi-directional topologies.
- Deployment model and data path. Self-hosted software (Qlik Replicate, Fivetran HVR, Tungsten, Debezium) keeps data inside your network but you operate it; SaaS platforms (Estuary, Striim Cloud, Confluent Cloud) reduce operations but data passes through the vendor unless you buy a private or BYOC deployment. Check this against your security and residency rules.
- Cost drivers. Pricing units vary more than in most categories: Monthly Active Rows counted per destination (Fivetran HVR), data volume plus connector instances (Estuary), task-hours plus data (Confluent), per-VPC packages (StreamSets) and quote-based licences scoped to endpoints (Qlik, GoldenGate, IBM). Model your table count, change rate and number of targets before comparing quotes.
- Vendor lifecycle. Confirm the product is still sold standalone and on which release line. SQL Data Sync retires on 30 September 2027, Arcion is no longer offered on its own, IBM now owns StreamSets and Confluent, and GoldenGate 23ai and later run only on the Microservices Architecture.