Compute Isolation Boundary
The line across which one domain's analytical workload cannot affect another's performance, cost attribution or access.
A shared data platform has to answer three separate isolation questions, and they have different answers. Performance: can one team's query slow another's? Cost: can spend be attributed to the team that caused it? Access: can one team read another's data?
Access is usually solved first because it is the one with an obvious owner. Performance and cost are frequently left shared, on the reasoning that a single large warehouse is cheaper — which is true until the first month-end when the finance close competes with a data science backfill.
The workable arrangement gives each domain its own compute, sized to its own pattern and tagged for its own cost, over shared storage with granular access control. Storage sharing is what avoids copies; compute separation is what avoids interference. Splitting the other way round — separate copies of the data, shared compute — is the arrangement to avoid, and it is what teams arrive at when each domain procures independently.