Skip to content

Bulkhead — Junior

At junior level, focus on this question:

How can a slow, unrelated dependency starve calls to a completely healthy dependency, if they share the same resource pool?

A shared thread pool with one slow dependency

flowchart LR Pool["Shared thread pool\n(20 threads total)"] --> CallA["Calls to Dependency A\n(slow, hanging)"] Pool --> CallB["Calls to Dependency B\n(healthy, fast)"] CallA --> AllThreads["All 20 threads eventually\nblocked waiting on\nDependency A"] AllThreads --> Starved["Calls to Dependency B\ncan't get ANY thread -\nstarved, even though B\nis perfectly healthy"]

If your application uses one shared thread pool for all outgoing calls, and Dependency A becomes slow (each call takes 30 seconds instead of 100ms), threads calling A pile up waiting on it — eventually consuming every thread in the shared pool. At that point, a request needing to call the completely healthy Dependency B has no thread available to even attempt that call — B's own health is irrelevant; the shared pool is exhausted.

The ship-compartment analogy

flowchart LR subgraph NoBulkhead["Ship with no internal walls"] Leak["One hull breach"] --> Sink["Water fills the ENTIRE\nship - it sinks"] end subgraph WithBulkhead["Ship with watertight compartments"] Leak2["One hull breach"] --> Contained["Water fills ONLY that\ncompartment - ship stays\nafloat"] end

A bulkhead in software is the same idea: partition a shared resource (thread pool, connection pool) into separate, isolated pools per dependency (or per criticality tier), so that one dependency's problems can only exhaust its own allocation, never spilling over to starve calls to unrelated, healthy dependencies.

🎓 Takeaway: the danger isn't the slow dependency itself — it's that a shared resource pool lets one dependency's problem become everyone's problem. Bulkheading contains the damage to exactly the dependency that's actually struggling.

Test yourself

  1. Why does a slow (not failed) dependency cause more resource exhaustion than a dependency that fails instantly with an error?
  2. In the shared-pool example, would a circuit breaker on Dependency A alone have prevented the thread-pool exhaustion, or is a separate fix needed?
  3. Why is "watertight compartments" a good mental model for what a bulkhead does to a shared resource pool?

Continue to middle.md.