Technology Participation in Enterprise Intelligence — browse technologies by architectural contribution.
Starburst and Zetaris as implementation contributors
Arqua architectural interpretation of selected public documentation, refreshed 17 October 2026. Not a benchmark, supplier certification or deployment recommendation.
A service-review question can require observations from one domain, asset relationships from another and applicable procedure information from a third. Federation can help combine those accounts without first rebuilding them in one central database. Its value is the useful composition it enables, not simply the number of connections.
This explanation examines Starburst Enterprise Platform and Zetaris's documented federated SQL and virtual-view facilities. They overlap in responsibility but have different execution and control arrangements. Neither is presented as the enterprise architecture, a preferred supplier or part of a mandatory combined stack. The selected accounts do not establish the capabilities of every edition, connector or installed release.
Begin with the question and participating information
Arqua's canonical definitions distinguish accepted representations, published products, formation of understanding and Context Qualification. Federation can implement selection and composition within Runtime Context Assembly. Qualification remains a distinct responsibility within RCA; a successful query is not, by itself, sufficient understanding or permission to act.
Define the subject, purpose, authorised population, relevant time and required claims before selecting an engine. Establish which sources and meanings are accepted for that use. A reusable view can implement part of a governed product, but saving SQL does not establish its owner, permitted uses or publication standing.
For example, the same local asset number can identify different assets in separate domains. Joining on that number alone can produce valid SQL and an invalid business relationship. Preserve namespaces, relationship meaning, applicable periods and cardinality. Federation should carry those decisions, not silently make them.
Starburst: reusable views and explicit access context
Starburst's view documentation describes a logical view as a stored query rather than stored data. The default DEFINER mode uses the view owner's permissions for underlying tables; INVOKER uses the querying user's. The current_user function identifies the invoker in either mode. These are legitimate alternative access arrangements, not interchangeable evidence of the backend identity.
Record the actual caller, selected roles, view security mode and source credentials where relevant. Authorised delegation may allow a consumer to use a view without direct base-table access. It does not automatically transfer the view owner's enterprise decision rights.
Available controls and configured controls are different evidence. Starburst's system access-control reference describes the unconfigured default as permitting operations except user impersonation and graceful shutdown; its separately selectable allow-all implementation permits all operations. Establish the effective system and connector controls for the selected deployment rather than infer restrictions from product capability alone.
The selected built-in access-control reference documents row filters and column masks. Materialized views do not inherit underlying column masks, so masks must be applied directly to them. The documentation also warns about joining masked columns. A transition from a logical view to a retained result therefore needs its own policy and interpretation review; the source controls must not simply be assumed to follow the copy.
Starburst: inspect the physical path and retained evidence
The pushdown reference makes execution dependent on connector and source support. Its generic same-catalog join condition is not a universal restriction across SEP: the Stargate documentation describes configurable cross-catalog pushdown where both catalogs support it. Inspect the selected plan and observed execution rather than infer placement from the SQL text. An optimiser estimate is not a measured cost.
Table scan redirection can transparently substitute an equivalent table in another catalog after source-catalog authentication and permission checks. A configured cache service can create and refresh those tables. This can be useful, but the unchanged logical query may now depend on a different physical copy and refresh history. Federation does not imply that no information is moved or retained.
The backend-service reference describes query history and configurable plan storage. Query-history persistence requires explicit enablement in the reviewed account. Such records can support reconstruction of execution; they do not automatically retain every result row, establish source authority or reproduce a cross-source snapshot.
These technical pages returned 481-e STS headers, while the newer indexed 482-e LTS release record describes subsequent changes, including CHAR/VARCHAR comparison behaviour. Direct retrieval of that release page failed. The references are therefore not one reconciled deployment baseline, and this article does not recommend the older build. Pin and check the supported version and connector configuration before implementation.
Zetaris: federation, AI Data Harness and governed data context
The Zetaris SQL Manual describes joins and views across multiple sources. A view can provide a reusable relational composition rather than merely expose disconnected tables. The manual also contains Lightning terminology and differing SQL examples; this explanation does not supply executable syntax for an unverified build.
Query Director describes selectable Spark and Presto engines, explicit query routing and user/role assignments. These are useful workload-placement controls. Choosing an engine does not demonstrate that every operator ran at a source or that the chosen route was cheapest.
Zetaris now publicly positions its AI Data Harness as a data layer for AI systems and agents working across distributed enterprise data. Current Zetaris Cloud material shows AI coding tools such as Cursor consuming live governed data context through Zetaris interfaces, while the Open Agent Hackathon separates Zetaris as the Data layer from the agent-building and model-execution layers.
This extends the implementation question beyond federated SQL alone. A data harness may contribute to source discovery, selection, composition and delivery of context to an AI consumer. That contribution is architecturally useful, but it is not the same as establishing that the selected context has accepted enterprise meaning, complete coverage, sufficient freshness or authority for a consequential action.
The Data Catalog documents browsable data sources, table and column metadata, metadata search, tags and metadata change history. Those facilities can support discovery and evidence about what participated. They do not by themselves determine which sources are authoritative for a particular purpose or whether a resulting context is sufficient for reliance.
The SQL Editor documents virtual-source queries, caching and materialisation, query history and a maximum-result-size control. Distinguish the virtual definition, retained data, query execution and result actually delivered. A displayed result must not be assumed to contain the complete population required by the question.
Roles and permissions distinguish interface privileges from access to data objects. The quality-and-exception example describes configured rules, record-level exceptions and daily reporting. These are substantive native contributions, not proof of synchronous quality checking on every query or identical enforcement through every engine, cache and client.
The dated 2.4.1.0 release note remains historical evidence, not a latest-release recommendation or a version binding for every current Cloud capability. Fix the actual product, version, interfaces and configuration before evaluating a production use. Starburst-specific controls are not attributed to Zetaris, or vice versa.
Data access and selection are not the end of the assurance chain. For an AI or agent use case, preserve the exact purpose, source identities, meanings, policy basis, freshness, missingness, provenance and selection conditions that produced the context. A successful data-harness response does not by itself establish qualified context, institutional authority or permission to act.
Compare the same obligation, not just matching rows
A fair comparison begins with the same subject, accepted meanings, authorised population, temporal question and required evidence. Separate what the provider offers, what the consumer requests, the bounded response and the assessment of that exchange. This is an architectural comparison discipline, not a claim that either supplier emits an Arqua-native contract.
For each requirement, identify evidence produced by the platform, an adapter or an accountable domain, and anything unavailable. Do not fabricate native fields or quietly lower the requirement to fit the weaker evidence path. Equally, a reviewed native control can be sufficient; another software layer is not automatically required.
Matching values do not establish equivalent identities, policies or source histories. Different results can be legitimate when consumers have different authorised populations. Compare cost and latency only after establishing comparable meaning, coverage and security. Include participating compute, source load, transferred bytes and retained intermediate data where measurable; disclose measurement gaps.
Keep time, completeness and failure visible
Separate the time the question concerns from when information was recorded, checked, cached and received. Two individually recent sources need not represent one coherent moment. A declared tolerance for source-time differences is not proof of an atomic snapshot.
An empty result, unmatched relationship, policy-filtered population, truncated delivery and failed source are different conditions. Preserve permitted evidence of those differences without revealing restricted record existence. A supported source observation may remain useful even when the broader conclusion lacks required evidence. Follow each engine and client's failure contract; this does not permit salvaging an invalid failed-query result.
Re-running unchanged SQL against changing sources creates a new result. Explaining an earlier decision may require retained result evidence, source revisions, view and mapping versions, identities and relevant checks. Protect that evidence according to its purpose rather than retain all sensitive rows indefinitely. Stored query history is not automatically accepted Institutional Memory.
Start with one read-only comparison
Use a bounded service-review question with observations, dependencies and procedure metadata. Establish the expected meanings and evidence needs, then compare each candidate implementation against an existing governed query or report. Introduce an ambiguous identifier, a stale retained copy, a restricted population and incomplete delivery. Evaluate which claims remain supportable, not merely whether the query finished.
This is a proposed assessment, not a report of supplier tests. No Starburst or Zetaris deployment, query, adapter or installed Guide was operated for this explanation. Public documentation has mixed retrieval ages, and no performance, residency or universal no-copy assurance is established. Wider AI, natural-language and unstructured capabilities are outside this selected scope, not presumed absent.
The objective is comparable, purpose-bound understanding across distributed sources—not to mistake one SQL interface, one successful query or matching rows for equivalent meaning, authority and evidence.