.svg){kind=logo}
A Private Knowledge Graph (PKG) is an environment-specific map of users, workloads, query patterns, ownership signals, Git context, and fix history on Snowflake or BigQuery. It is the ownership substrate under APEX: without it, Trace can still show an object, but Assign cannot name a party you will trust. When ownership is not knowable, it says needs owner instead of guessing.
Key takeaways
Why generic models cannot assign your issues
Industry-average “right-size this warehouse” advice ignores your coding patterns, your BI cluster’s shared schedule, and which engineer fixed this class of regression two quarters ago. Wrong assignment is worse than no assignment: you accuse the wrong team of wasting money. The PKG exists so routing uses your identity map, your query signatures, and your fix history.
That is also why workload governance is a third-party layer. Snowflake’s native views are necessary telemetry. They are not a memory of who owns what across billing cycles.
What the graph maps
- Users, roles, and service accounts to teams and, when declared, cost centres — on Snowflake and on BigQuery
- Snowflake workloads: recurring queries, warehouses, pipelines, dashboards, Cortex functions
- BigQuery objects: jobs, datasets, reservations, slots, and scan patterns — not a 200+ BigQuery catalog claim
- Cost signatures: idle, spill, scan, refresh storms, storage retention
- Ownership evidence: tags, roles, usage hints — with confidence that never promotes a guess to a personal accusation. If it is not knowable, the graph says needs owner.
- Fix history: what changed, when, and whether the waste stopped
On Snowflake, Anavsan’s PKG ingests 200+ signals across queries, warehouses, storage, Cortex, serverless compute, and related logs. BigQuery uses the same ownership substrate from INFORMATION_SCHEMA-class metadata — we do not invent a 200+ BigQuery signal catalog. Continuous sweeps keep the graph current; assignment quality improves as history accumulates (typically weeks, not minutes).
What it is not
| Not this | PKG | |
|---|---|---|
| Data graph | A catalog of customer records | A map of metadata about how you use Snowflake or BigQuery |
| Context graph / AutoApply | A graph built so an agent can AutoApply changes | Ownership substrate. It names a party or says needs owner. It does not AutoApply. |
| LLM memory | Chat history pasted into a prompt | Structured links among objects, people, and outcomes |
| CMDB clone | A static inventory you maintain by hand | Assembled from platform metadata and updated on a cadence |
How it is used
Trace uses the graph to attach a workload to credits. Assign uses it to pick a channel: high-confidence declared owners can reach a person; inferred ownership stays at team level. Prove uses prior similar fixes as context. Enforce uses the trail so a returning pattern is not a brand-new alert. Product behavior is documented on APEX.
Related terms
APEX
The engine that builds and uses the PKG.
Read →Trace → Assign → Enforce → Prove
The loop the graph makes specific.
Read →Cost Accountability
The outcome: a named owner, not a guessed one.
Read →Enforcement Desk
Where PKG-backed findings show up as rows.
Read →