A reservation could be accepted while its hospital database was offline. That was the product constraint behind the CDC work I handled at SmartDoctor: retain integration intent, then deliver it across a boundary with a different availability schedule.

Kafka and Debezium supplied the event path. Strimzi made their deployment part of the Kubernetes operations model already used by the application services.

Capture recorded integration work

The retained MySQL connector is scoped to inserts into the reservation platform’s outbox. Its consumer reads those events and dispatches hospital-scoped SQS messages or the specified API work. It marks the handled entry read and acknowledges the Kafka record afterward.

The hospital-side consumer then applies the reservation under the appropriate hospital context. Classified offline-database failures take a different retry path from other processing errors.

This is outbox CDC. It does not mean Debezium was watching every hospital’s SQL Server database, and it does not make an external API or SQS effect atomic with the consumer acknowledgment.

Why Strimzi fit the operating model

The charts express Kafka, Kafka Connect, connectors, broker/controller node pools and environment configuration as Kubernetes resources. That let the event infrastructure live alongside the application’s Helm and GitOps configuration rather than become an unrelated set of manually managed processes.

The division of responsibility was useful:

ResourceResponsibility in the pipeline
Kafka / node poolsBroker and controller configuration, placement and storage
KafkaConnectWorker image, resources and internal configuration/offset/status topics
KafkaConnectorMySQL source, included outbox table and capture settings
Application consumerDispatch, acknowledgment, retries and receiving-side correctness

The operator did not remove the need to understand broker storage, replication, connector compatibility or failure recovery. Two Connect replicas are not enough to establish high availability when internal topics or the rest of the pipeline have weaker redundancy.

Connector configuration is part of the application contract

The reviewed connector selects outbox inserts and skips update/delete operations. That matches the consumer’s interest in newly recorded work. The application marks handled entries read separately; that update is not a new integration command.

The Connect image includes the connector plugin, and the deployment specifies configuration, status and source-offset storage topics. Environment-specific settings still have to agree with the application’s topic and consumer configuration.

Historical capture from the Strimzi deployment investigation

Acknowledgment does not eliminate replay

Suppose a consumer sends a reservation message successfully, then fails before recording completion or acknowledging the Kafka record. Reading the record again can repeat the dispatch.

The producer uses hospital message groups and a create-message deduplication identity. Receiving operations still need a durable correctness contract; queue deduplication is not permanent exactly-once database execution.

This separation made it possible to reason about accepted bookings, retained integration intent, dispatched messages and hospital-side completion independently. The offline-database case explains the connection and retry behavior.

Recover the source position deliberately

I also investigated connector failures involving a recorded MySQL binlog position that was no longer available. Recovery depends on source-log retention, the connector’s source-offset record, schema history and its snapshot mode.

A Kafka sink-consumer group offset and a Kafka Connect source connector’s stored MySQL position are different state. Resetting a consumer group is not a general repair for a missing source binlog. Skipping ahead can also leave a delivery gap that requires reconciliation.

The operational investigation therefore needs to preserve the error, inspect the source position and retained logs, and determine a supported snapshot or reconciliation path before changing offsets. This article does not claim a newly reproduced recovery result.

The useful design outcome was a Kubernetes-managed pipeline for committed outbox work, with the application boundary kept visible. Strimzi organized deployment; it did not decide when a hospital operation was complete.

References: Strimzi deployment and operations, Debezium MySQL connector.