Bulkhead pattern: isolating failures in multi-tenant systems

We recently had an incident where a large bulk import job from one of our tenants consumed all available resources, impacting the performance and availability for every other tenant on that service. It highlighted a critical need for better failure isolation. We're looking into implementing the Bulkhead pattern. What are the most effective ways you've implemented bulkheads in a multi-tenant system? We're considering separate connection pools per tenant, aggressive rate limiting, or even separate compute instances for certain high-risk tenants or operations. How do you balance the level of isolation with the efficiency of resource utilization, especially when you have thousands of tenants? Any specific technologies or patterns that have worked well for you?