ARCH-ONT-001 — Align the Cube execution topology
| Field | Value |
|---|---|
| Status | Open finding; deployment and source verified |
| Severity | High |
| Components | sparqd/platform-stacks, Cube, CubeStore, Ontology backend |
| Associated issues | Not opened pending finding review |
Concrete observed case
The deployment has one Cube API, one CubeStore router, and two CubeStore workers. Cube uses the in-memory cache/queue driver; refresh worker and scheduled refresh are disabled; generated models define no pre-aggregations. CubeStore workers therefore do not accelerate these raw Trino queries. AliCloud affinity places the router and workers on one node, and no Cube component has an HPA or disruption budget.
Expected condition
Deployment matches one declared mode:
- minimal raw-query mode right-sizes unused CubeStore/refresh capacity and scales Cube API and Trino from measured demand; or
- accelerated mode uses the production distributed queue/cache, separate refresh capacity, secure pre-aggregations, durable storage, and failure-domain separation.
Why it matters
Current fixed resources add cost without raw-query capacity, while one API, in-memory coordination, and same-node placement retain availability limits.
Existing controls and workarounds
Replica counts exist in lower-level KCL modules, but the flavor does not expose all of them. A fixed topology is acceptable only with a measured low-load workspace envelope and recovery expectation.
Alternatives considered
- Recommended now: explicitly support and right-size minimal raw-query mode.
- Add accelerated mode after pre-aggregation freshness and Ranger user isolation are proven.
- Add CubeStore workers alone — rejected because it does not speed raw source queries in the current configuration.
Component involvement
- Platform stacks exposes coherent mode, sizing, placement, and durability.
- Ontology/Cube generation adds pre-aggregations under a reviewed contract.
- Operations measures queues, cache hits, refresh demand, and failures.
Reproduction and verification
Follow ONT-S020.
Resolution conditions
Close when the selected mode is documented, unused capacity is removed or made functional, and representative load/failure tests establish the envelope.
Validation and blockers
Topology and configuration are live/source-confirmed. Workload demand, pre-aggregation design, and recovery objectives remain undecided.