The mechanism by which an overloaded downstream signals upstream to slow down, instead of dropping requests silently.
Backpressure is a flow-control pattern where a downstream component, when it's nearing saturation, signals upstream components to send less traffic rather than silently dropping or queueing it. HTTP 429 (Too Many Requests), reactive-streams Subscription.request(n), and gRPC flow-control windows are all backpressure mechanisms. The point is to make load shedding explicit and observable instead of letting queues grow until they fail in unpredictable ways.
Without backpressure, an overloaded service either drops requests (silently bad) or queues them indefinitely until memory exhausts and the process crashes (loudly bad). With backpressure, the upstream sees the slow-down signal and can either retry with backoff, route to a different instance, or surface the slowness to the user. Most cascading-failure outages are missing-backpressure stories.
See the part of the platform that handles backpressure in production.