2.4 Data Types, the Data Supply Chain, and Interoperability

Version Note

Everything in this section belongs to the exam guide effective August 12, 2026, which expands sub-topic 2.1 with data types, the data supply chain, and interoperability. Candidates sitting the current exam through August 11 already have the value-of-data material in Section 4.1 and can treat this as background.

Classifying Data by Source: First, Second, and Third Party

The party framing is about relationship and consent, not about volume or quality.

The exam's business point: first-party data is the differentiator. Any competitor can buy the same third-party segment, so a model trained on purchased data confers no advantage. A model trained on your own customers' behaviour cannot be replicated. This is also why privacy regulation and browser changes that restricted third-party tracking pushed organizations to invest in first-party data collection.

Classifying Data by Structure

Semi-structured is the one candidates miss. It is not a halfway house between the other two in quality; it is a genuinely different shape. A JSON document is self-describing - each value is labelled by its key - so software can parse it without a schema declared up front, yet different documents in the same collection may carry different fields. That flexibility is exactly why event streams and application logs use it, and why document databases such as Firestore exist.

A useful test for the exam: ask whether the data carries its own labels. Rows in a table get their meaning from the schema, so structured. A photograph carries no labels at all, so unstructured. A JSON payload labels its own fields, so semi-structured.

The Data Supply Chain

The updated guide replaces the current guide's "data value chain" with a six-stage data supply chain. The stages are more granular at the front end.

Genesis is the stage worth pausing on, because it is the one the older framing omitted. Data quality is decided at genesis: a field the source application never captured cannot be reconstructed downstream, and a timestamp recorded in local time without a zone can never be safely compared across regions. Organizations that treat data as a product instrument the source rather than patching the pipeline.

And as with the older value chain, value is realised at activation. Every stage before it is cost. A supply chain that ends at a dashboard nobody acts on has produced nothing.

Data Silos, Openness, and Vendor Lock-In

A data silo is a store of data controlled by one team or system and not readily usable by the rest of the organization. Silos are rarely created deliberately - they accumulate as each department buys the tool it needs.

The damage compounds along the supply chain:

• Conflicting answers. Finance, sales, and support each compute revenue from their own copy, and no number is authoritative.
• Invisible data. A model can only learn from data it can reach; a silo is training data that does not exist as far as the model is concerned.
• Duplicated cost. The same record is stored, processed, and secured several times.
• Governance gaps. A silo nobody catalogued is a silo nobody is applying retention or access policy to.

The updated guide pairs silos with openness and interoperability as the remedy, and the argument is worth stating precisely because it is a business argument, not a technical preference:

1. Open formats — storing data in Parquet, Avro, or Iceberg rather than a proprietary internal format means more than one engine can read it, so choosing a new analytics tool does not require migrating the data.
2. Open engines and standards — PostgreSQL compatibility, standard SQL, Apache Spark, and Apache Beam mean the skills and the code transfer.
3. Products that run off the provider's own cloud — BigQuery Omni queries data left in AWS S3 or Azure Blob Storage; AlloyDB Omni runs the AlloyDB engine outside Google Cloud. Data does not have to move to be useful.

Vendor lock-in is the condition where the cost of leaving a provider exceeds the benefit of leaving, so the provider's terms become effectively unilateral. Interoperability does not eliminate lock-in - it bounds it, keeping the exit cost known and finite. That is the honest framing, and it is the one an exam scenario about "negotiating leverage" or "regulatory requirement to demonstrate portability" is looking for.

Governance Ties It Together

Data governance - the policies and controls that make data trustworthy, secure, and responsibly used - is what stops a supply chain from silting up into silos. Applied at the storage and processing layer rather than per application, governance gives you one catalogue of what exists, one lineage graph of where each field came from, one access model, and one retention policy. Without it the six stages still run, but nobody can prove the output is correct - and an unprovable number is not a decision-grade number.