Dedicate Qdrant + CNPG database to octopus, isolated from engram
Date: 2026-06-06 Status: Accepted Decision: hl-msh Deciders: Sean Brandt
Context
Section titled “Context”The cluster already runs a Qdrant instance (engram/agent-memory) and CNPG databases (authentik, grafana, tandoor, litellm). Octopus requires both Qdrant (vector store) and Postgres (pg-boss queue + application data). The question is whether Octopus shares these with existing workloads or gets dedicated instances.
Decision
Section titled “Decision”Deploy a dedicated qdrant/qdrant Deployment with a Longhorn-encrypted PVC in the octopus namespace, and provision a dedicated CNPG Database CR (octopus) on the existing main cluster with an additive managed.roles entry. Octopus shares no Qdrant or Postgres resources with engram or other workloads.
Rationale
Section titled “Rationale”- Isolation eliminates cross-workload blast radius: Octopus can be rolled back (scale-to-0) without affecting engram’s memory or any other database.
- The
EMBEDDING_DIM=1024Qdrant collection schema is Octopus-specific; sharing engram’s Qdrant would require collection-level namespace management that upstream Octopus does not support. - The CNPG
DatabaseCR +managed.rolespattern is already established for tandoor/authentik/grafana/litellm — Octopus follows it without novel infrastructure.
Alternatives Considered
Section titled “Alternatives Considered”- Share the existing engram Qdrant instance — no extra resource cost, but Octopus’s 7 collections would co-locate with engram’s; any Qdrant operator error/upgrade affects both, plus collection-name collision risk. The hl-cy3 dual-backup strategy was designed for engram’s single-tenant use.
- External managed Postgres (not CNPG) — reduces stateful footprint, but contradicts the cluster’s CNPG-for-all-Postgres pattern and adds an external dependency for an intended fully-in-cluster workload.
Consequences
Section titled “Consequences”- Positive: Octopus rollback is fully reversible with no effect on other workloads; engram’s hl-cy3 backup strategy is unaffected; the
managed.rolesaddition is additive-only. - Negative: a second Qdrant Deployment increases cluster memory (~256–512 MiB idle) and adds an instance to monitor/upgrade independently.
- Neutral: the dedicated Qdrant does not inherit engram’s hl-cy3 dual-backup automatically; an Octopus Qdrant backup decision is deferred.