Change an owner. Predict the outcome. Step through a collection and explain what survives.
A deterministic teaching model for CoreCLR concepts. It does not inspect your device, run the real GC, show real addresses or predict collection timings. Unity and Mono/Wasm use different collectors.
Select an object to see its retaining path.
This sketch connects the diagram to a language construct. Run the full companion program in the linked lesson; the model is not its heap dump.
Open the detailed lesson ↗Throughput capacity = workers × 1,000 / service time. Expected active requests = min(workers, rounded-up arrival rate × service time). Backlog grows at the positive difference between arrivals and capacity, capped by queue size, over ten seconds.
Payload live set = (active + queued) × payload size × simultaneous copies. Pool, managed baseline and native allowance are separate assumptions. The pool value excludes active payload buffers to avoid double counting. Fractional rates are a fluid approximation; this is not a queueing-theory latency or p99 predictor.
A full queue needs an explicit wait/reject/drop contract. Waiting producers may retain payloads outside the channel; the capacity of a channel alone does not bound those callers. Acquire admission before allocating, bound waiting callers, or include their retained memory. Slow clients and retries can extend ownership too.
Use the diagram to form a hypothesis. Prove the actual retaining path with a matching-runtime dump, measure allocation rate and surviving bytes separately, then rerun the same workload after a change. A falling live set does not imply an immediate fall in process resident memory.