Phase 1 Investigation Backlog

On this page

Summary

Single backlog for active investigations that feed architecture and ADR updates.

The previous investigation notes are consolidated here to avoid duplication and preserve open decision points.

Domain: OpenShift
Status: investigating
Focus: Investigate ACM/MCE hub topologies for a centralized operating model: a single ACM hub with multiple Red Hat Advanced Cluster Management instances (multi-engine hub) versus multiple ACM hub clusters (one MCE per hub). Multi-engine hub supportability remains unconfirmed until validated against vendor guidance and baseline comparisons.

Domain: OpenShift
Status: investigating
Focus: Define internal LLMeval benchmarks, LMjudge rubrics, and release-gate integration for coding-agent models that accelerate internal DevOps.
LLMeval: internal automated benchmark suite used to score coding-agent quality across representative DevOps tasks.
LMjudge: rubric-based scoring process (model-assisted and/or human-reviewed) used to judge quality, safety, and policy compliance.
Data controls: evaluation datasets and prompts follow data-classification policy, with redaction and anti-leakage checks before pipeline ingestion.
Audit controls: evaluation outputs, threshold decisions, and release-gate overrides are retained with traceable approval records.

Domain: Storage
Status: investigating
Focus: Document how Ceph is consumed as the platform’s primary storage backend, including ODF internal mode versus external RHCS-backed mode assumptions, pool layout, tenant isolation, access patterns, and operational constraints.

Domain: OpenShift
Status: closed
Focus: Investigation into using one Clusters as a Service cluster per site (one cluster per datacenter) to host shared services.

When does Multicluster Global Hub add value beyond observability/reporting, and what operational overhead does it add?
What is the supported approach for migrating hosted clusters between MCE instances or hub clusters in current releases?
What are the scale limits for a single ACM hub before a multi-hub model becomes necessary?

What is the minimum coding-agent evaluation set needed for every release?
Which evaluation thresholds should block promotions versus raise alerts?
How do we handle dataset refresh cycles and rubric updates?
Which data classes are allowed in evaluation datasets, and what anti-leakage controls are mandatory?
What is the required human-override and governance process when a blocking evaluation gate is bypassed?