A Dataset is a collection of data objects (files or blobs) and associated metadata objects. And are identified by a name and a version. These values are used when publishing or subscribing to availability events, and for internal naming of metadata resources.

Data objects are the actual data, either consumed or produced, externally or by arc workloads within the framework.

Typically, data objects are organized by partitions, like year=2023/host=abc-001. Frequently partitions are not named, 2023/abc-001, but the data is still organized for use.

Datasets are also organized using a lot identifier.

The metadata objects are manifest files that list the locations of objects arrived within a given lot interval.

If a hundred files arrived in an S3 bucket between 0:0:0 (inclusive) and 0:5:00 (exclusive), the manifest could be labeled as having lot id 20230101PT5M000.

Manifest files have state. They may be complete, partial, empty, or removed.

The empty state is important. A workload may not produce any data, and this outcome may be important to downstream listeners.

When a workload fails and writes partial data, this data needs to be clean up or leveraged in a new retry attempt for the workload. Marking a manifest as being removed confirms duplicate data isn’t interleaved in the dataset and provides some accounting for the failures.

See the ADR arc-state-and-data-metadata for more details on manifest state.