Collection-per-tenant creates overhead that scales poorly
For large multitenant systems, a shared collection usually scales operationally better because each collection has its own internal indexes, storage structures, lifecycle, monitoring, backup, and maintenance concerns. Thousands of small collections can create significant aggregate overhead even when many tenants have little data. A shared collection with tenant-partitioned payloads lets the system manage fewer vector indexes while filters provide logical partitioning. The trade-off is shared-resource contention and reliance on correct tenant scoping. I would still use dedicated collections or deployments for tenants requiring regulatory isolation, separate residency, stronger performance guarantees, or independent operational control.
Fewer collections reduce aggregate index and lifecycle overhead and simplify monitoring, backups, upgrades, and capacity management.
Trade-off: shared collections reduce isolation between tenants at the infrastructure-resource level, so noisy tenants need monitoring and potentially dedicated capacity.
One collection per tenant is not automatically more secure; the actual authorization and isolation model is what matters.
For highly skewed tenant sizes, a tiered architecture can move exceptional tenants to dedicated collections or deployments while keeping ordinary tenants shared.
Your team proposes 5,000 collections for 5,000 customers. What questions would you ask before accepting that design?
A tenant has only 100 vectors but gets its own collection. What kinds of operational overhead might result?
You inherit one collection per tenant and collection count grows by hundreds each month. How would you evaluate consolidation?
A shared collection is easier to operate, but one tenant generates most query traffic. What mitigation would you investigate?
How would you migrate 10,000 tenant-specific collections into a shared collection without losing tenant ownership information?
Your largest 1% of tenants account for most vectors and traffic. What collection and deployment topology would you choose?
Design isolation tiers for thousands of small tenants, hundreds of medium tenants, and a few regulated enterprise tenants.
What metrics and failure-mode analysis would you use to decide whether shared multitenancy remains viable as tenant and vector counts grow?