Skip to content

Latest commit

 

History

History
86 lines (57 loc) · 4.8 KB

concepts.md

File metadata and controls

86 lines (57 loc) · 4.8 KB
Raw

Concepts

The Arcane platform consists of the following components, which can be installed using Helm charts:

  • Arcane Operator that manages the lifecycle of the streaming jobs
  • Arcane Stream Plugins: REST API, Microsoft SQL Server, Microsoft CDM change feed, and more to be added in the future.

During the installation Arcane Operator deploys the following CRDs:

  • StreamClass, which is used to define configuration of the stream plugin.
  • StreamingJobTemplate, which is used to generate a Job object for the Kubernetes.

The configuration of each individual streaming job is managed by the custom resources installed by the stream plugin Helm Chart. These resources referred to as StreamDefinitions in the text below.

During the stream plugin installation process the helm chart installs the StreamClass object that contains the StreamDefinition API group, version and plural name.

Also, the plugin Helm chart may install the optional StreamingJobTemplate object, that contains the Kubernetes Job definition that the Operator uses to generate streaming jobs.

Stream reliability

At the beginning of the streaming job, it fetches a small volume of historical data to ensure that the stream fetches all data, even in the case that the streaming job was restarted for some reason. The depth of the historical data is controlled by the lookBackInterval parameter in the StreamDefinition object.

Backfilling and schema evolution

Currently, supported stream plugins use a polling mechanism to fetch data from the source. If the source supports schema evolution and the source data schema has changed, the stream plugin completes and requests the Arcane Operator to backfill the data.

The data backfill process is implemented by the Arcane Operator by creating a new job with an (potentially) different job template and providing the STREAMCONTEXT__BACKFILL environment variable to the streaming job.

After the backfill is completed, Arcane Operator resumes the streaming job from the point where it was stopped.

Secret Management for streaming jobs

Arcane Operator mounts the Kubernetes secrets to the streaming job pods as environment variables. Each StreamDefinition object may contain a reference to a Kubernetes secret that should be serializable to the V1SecretEnvSource object.

The list of the StreamDefinition fields that should be serialized to a V1SecretEnvSource object defined in the StreamClass object managed by the stream plugin Helm chart.

Communication between Arcane operator and Stream Plugin

The communication between the stream runner and the Arcane operator is done in two ways:

  1. The Arcane Operator creates a streaming job in the Kubernetes cluster and provides the stream configuration through the environment variables.

  2. The stream runner can finish with the Success status and in this case, it will be restarted in the streaming mode.

  3. The stream runner can finish with the Failed status and in this case, the Operator will create the arcane.streaming.sneaksanddata.com/state: crash-loop annotation on the stream definition and stop processing the SD events until operator removes this annotation manually.

  4. The stream runner may annotate the streaming job with the arcane.streaming.sneaksanddata.com/state: schema-mismatch annotation and exit with the Success status. In this case, the operator restarts the stream with the backfill mode and removes arcane/state: schema-mismatch annotation from the stream definition object.

Streaming Job lifecycle

The streaming job always starts in backfill mode when added. This ensures that the streaming job fetches all the data available from the source on the first run.

When the backfill is completed, the streaming job restarts in the streaming mode.

To suspend the streaming job temporarily, the streaming job may be annotated with the arcane/state: suspended.

Note

If the stream definition was created in a suspended state and unsuspended later, the streaming job will start in streaming mode, not in backfill mode. This behavior is intended to prevent unnecessary backfilling of data in case of migration of the streaming job from one cluster to another, or any other activity that requires the stream definition to be recreated.

If the streaming job finished in the success state, the streaming job will be restarted in the streaming mode. If the streaming job is finished in the failed state, the streaming job will be annotated with the arcane/state: crash-loop and will not be restarted until the annotation is removed. This behavior is intended to prevent misconfiguration of the streaming job that may lead to the infinite restart loop.

The ability to change the lifecycle annotations to be implemented in the future.