Definition
Backpressure is a flow-control mechanism in which a component that is falling behind signals upstream producers to slow down or stop, rather than accepting work it cannot process. The signal can be explicit (a 429 response, a blocked write, a paused consumer) or implicit (a bounded queue that refuses new items once full).
The alternative to backpressure is unbounded buffering, and unbounded buffering is how distributed systems fail badly rather than gracefully. A queue with no ceiling absorbs the overload as growing latency and memory until the process is killed — at which point everything in that buffer is lost at once. A bounded queue that rejects work sheds load early, keeps latency predictable for the requests it does accept, and pushes the decision about what to drop up to a layer that has the context to choose well.
How it comes up in interviews
Interviewers probe this by overloading your design: "the consumer slows to half throughput — what happens?" The weak answer is "the queue absorbs it", which just moves the failure. Strong answers bound every queue, name what happens at the boundary (reject, shed, spill to storage, block the producer), and explain who is told about it.