This content is currently locked.

Your current Infotech subscription does not include access to this content. Contact your account representative to gain access to Premium SoftwareReviews.

Contact Your Representative
Or Call Us:
+1-888-670-8889 (US/CAN) or
+1-703-340-1171 (International)

Change Data Capture (CDC) Implementation – Selected Vendors Overview

Technology Note By: Igor Ikonnikov, Infotech

CDC keeps downstream systems synchronized with changes in operational data. Vendors implement it through database logs, native change interfaces, application events, or comparison of successive states. These approaches differ in the changes they preserve, the load they impose, and the recovery guarantees they can support.

CDC should be evaluated as an end-to-end service. A capture connector is only one component. Initialization, durable delivery, target application, history management, and operational recovery determine whether the result can be trusted. Metadata strengthens that service by explaining each change and its meaning but does not replace it.

The main capture methods

Method Implementation and implications

Log-based capture

Reads transaction logs directly or through replication interfaces. Can capture committed inserts, updates, and deletes without repeatedly scanning source tables.

Native change tables or APIs

Consumes database-maintained change records or changed keys. SQL Server CDC provides change history, and Change Tracking identifies changed records with less historical detail.

Application events or triggers

Publishes application changes or writes them into logging tables. Coverage depends on configured objects and write paths. A durable event stream can support high throughput.

Watermark polling

Selects rows using a reliable timestamp, ID, or version. Usually requires a separate mechanism for hard deletes and cannot recover unseen intermediate states.

Snapshot comparison

Compares keys, hashes, or complete records between observations. Useful when capture access is unavailable. Differences between snapshots are not a full transaction history.

Separate data changes from schema changes. Capturing an update or delete does not prove that a connector can safely propagate a column rename, type change, or dropped field. These behaviors require explicit compatibility rules.

Implementation in major data platforms

Platforms increasingly combine managed ingestion with downstream incremental processing. The same CDC label can refer to either function. Verify the exact source, destination, deployment, and connector release state.

Platform Capture and ingestion Downstream behavior

Microsoft Fabric

Database Mirroring replicates supported operational sources into OneLake. Data Factory Copy Job supports CDC and watermark-based incremental movement for eligible connectors.

Mirroring maintains an analytical replica. Transformations and historical models remain separate design decisions.

Azure Data Factory

Native CDC in mapping data flows consumes supported source change interfaces. Watermark pipelines and the CDC Factory resource provide other patterns.

Apply insert, update, upsert, and delete operations according to source and sink support. Scheduled extraction need not be continuous streaming.

Databricks

Lakeflow Connect database connectors use an ingestion gateway, staging storage, and managed ingestion pipelines. Query-based connectors are a separate option.

AUTO CDC applies a supplied feed as current state or SCD Type 2 history. AUTO CDC FROM SNAPSHOT compares snapshots. Delta CDF exposes Delta table changes.

Snowflake

Openflow database connectors capture external source changes. The SQL Server CDC connector consumes SQL Server native CDC.

Streams expose changes to tables already in Snowflake. Tasks and SQL can apply them downstream. A Stream stores an offset and is not a permanent source-event archive.

AWS

Database Migration Service supports full load plus CDC and CDC-only tasks using source-specific log interfaces.

Target application depends on endpoint and task settings. CDC latency is variable.

Google Cloud

Datastream provides managed CDC and backfill into BigQuery or Cloud Storage. Dataflow supports additional processing and delivery patterns.

Choose direct replication or a processing pipeline according to target and transformation requirements.

Capture and history are separate capabilities

SCD Type 1 updates current state, while SCD Type 2 retains successive analytical versions. Neither automatically means every source mutation is retained. If the upstream connector supplies only net changes, a downstream history feature cannot reconstruct intermediate values that never arrived.

Implementation in integration and ETL tools

Dedicated replication tools emphasize capture and recovery across heterogeneous systems. ELT tools often add managed loading and schema handling. Transformation tools operate on changes already available in the analytical platform.

Tool Implementation What to verify

Oracle GoldenGate

Extract captures supported source changes, trail files buffer and transport them, and Replicat applies changes to targets.

Source and target support, transaction handling, restart checkpoints, and topology.

Qlik Replicate

Log-based capture with full-load initialization, transaction buffering, target apply, and an option to store changes.

Whether the output is a maintained replica or a retained change dataset.

Informatica

Cloud ingestion and replication services and PowerExchange offer source-specific capture, including enterprise and mainframe integration.

Exact product, connector, capture method, and supported deployment.

IBM Data Replication

CDC Replication captures source changes and can deliver them to DataStage through supported integration paths.

Distinguish transaction-log capture from a DataStage job that compares datasets.

Fivetran

Database connectors use native CDC or proprietary incremental methods. SQL Server options include CDC, Change Tracking, and Teleport comparison.

Full history versus latest-state synchronization. Delete handling and resync behavior.

Airbyte

Supported CDC connectors consume database changes and retain a log position. Its documented orchestration uses scheduled syncs.

A log-based source does not necessarily produce continuously visible target changes.

Matillion

A CDC agent captures source changes, buffers them, and delivers partitioned files into cloud storage.

Capture, file delivery, loading, and application at target each affects freshness.

Debezium

Database connectors capture changes, commonly through Kafka Connect; Server and embedded Engine are other deployment models.

Kafka provides transport and retention when used, and the connector performs database capture.

dbt

Incremental models process available data. Snapshots compare observed states to maintain analytical history.

Snapshots cannot recover source changes that occurred between observations.

Extract-transform-load (ETL) and extract-load-transform (ELT) describe where transformation occurs rather than how changes are detected. Both can consume CDC. Apply required security filtering before data crosses its permitted boundary, even when most transformations occur after loading.

In ETL, transformation happens before load, so a schema change (a column added, removed, or redefined) can break the pipeline or silently drop fields. ELT lands the data first and transforms it afterward, which is faster and tolerates schema drift better. Microsoft's move from Azure Data Factory to Fabric Data Factory is a commonly cited example. Verify the exact scope of that change before quoting it.

The target needs its own change policy. Capturing a change is only half the design. Teams must also decide whether to update in place, keep a shadow record, or keep only a change log. A cost-aware approach, for example, keeps only the latest state in the hot layer and pushes change history to cold storage for audit.

SQL Server and PostgreSQL offer different capture interfaces

Both engines support log-derived change capture. PostgreSQL logical decoding converts write-ahead log information into a consumable change stream. SQL Server native CDC asynchronously reads the transaction log and populates relational change tables.

CDC assessment SQL Server PostgreSQL

Strengths

Native CDC exposes row changes and sequence metadata through SQL-accessible tables and functions. Familiar to Microsoft operations teams.

Logical decoding and replication slots support independent consumers. A broad connector ecosystem uses these interfaces.

Weaknesses

Capture tables and cleanup create operational work. Source deployment, permissions, edition, and schema behavior need checking.

Slow consumers can retain WAL and consume storage. Replica identity, slot lifecycle, and version-specific failover behavior require attention.

Opportunities

Deliver incremental feeds from an existing estate without making database migration a prerequisite.

Use a consistent event contract across PostgreSQL deployments and exploit open tooling where operationally appropriate.

Threats

A consumer falling behind retained change history may need reinitialization. Schema or failover changes can disrupt capture.

Slot invalidation or unavailable WAL can require reinitialization. Restart can resend records, requiring duplicate-safe consumers.

This is an architectural SWOT, not a benchmark. Compare capture overhead, commit-to-target latency, sustained throughput, recovery time, storage amplification, licensing, and operating effort using the same workload and failure scenarios. Migration cost and application compatibility belong in the database decision.

Technical distinctions that prevent misleading comparisons

Database logs, table-format metadata, and business metadata serve different purposes. A database log supports recovery and replication; Delta or Iceberg metadata describes analytical table state and evolution; and a catalog or semantic layer explains business meaning and governs use. These capabilities complement one another.

Delta Lake provides its own change data feed. Unity Catalog is not a prerequisite for the open-source format to represent row-level changes. Snowflake supports native tables as well as Iceberg, and its CDC story is broader than Iceberg metadata.

A database or lakehouse does not automatically retain every change forever. Retention, cleanup, checkpointing, and consumer behavior determine recoverability. PostgreSQL documentation explicitly describes possible resends after restart and the storage consequences of retained replication slots.

What buyers and CIOs should evaluate

Require a demonstration of the exact source-to-target path. A connector listing establishes neither end-to-end correctness nor a freshness guarantee. Use a production-like workload and include failure tests before agreeing to service commitments.

Evaluation area Evidence to request

Completeness

Consistent initial load plus concurrent changes; inserts, updates, hard deletes, key changes, and documented exclusions.

Ordering and application

Stable keys and source sequence positions; duplicate-safe replay; defined behavior for late events and multitable transactions.

Recovery

Checkpoint durability, retention window, outage replay, source failover, reconciliation, and a controlled reinitialization procedure.

Schema evolution

Behavior for added or dropped columns, renames, type changes, new tables, and versioned consumer contracts.

Governance

Tenant isolation, authorization at delivery, masking, lineage, ownership, data definitions, and clear rules for downstream copies.

Economics

Capture and apply compute, storage, network egress, destinations, replays, initial loads, and support effort under peak workloads.

Compare pricing units against the workload

Row counts, events, and bytes measure different things. Fivetran uses monthly active rows, with a changed row generally counted once per month in its standard model. Infor Stream Pipelines measures daily processed record variations and includes initial-load and batch events. These meters produce different costs for frequently updated records. Check product-specific exceptions and contractual entitlements.

A stream subscription offers destination flexibility but assigns consumers responsibility for applying changes, handling schemas, and recovering their targets. Managed replication reduces that consumer burden and adds connector and support costs. A fair comparison includes both provider charges and customer operating effort.

Our Take

Prefer log-derived CDC or a durable application change feed when complete change capture and low latency matter. Use polling or snapshot comparison where source access or workload economics justify their reduced historical fidelity. Throughput alone does not disqualify application events; durability, coverage, and backpressure handling are the deciding factors.

Treat convergence of operational and analytical platforms as a way to simplify selected paths. CDC remains necessary when data must cross systems, clouds, or ownership boundaries. Invest in capture reliability alongside business metadata so downstream analytics and AI can determine both what changed and what the data means.

Want to Know More?

Latest Technology Notes

All Technology Notes
Over 100 analysts waiting to take your call right now: +1 (703) 340 1171