Back-Pressure — Senior¶

At senior level, focus on this question:

How must back-pressure propagate across a chain of multiple services, not just one producer-consumer pair?

Prerequisite: middle.md.

A back-pressure signal that stops at the first hop is incomplete¶

flowchart LR A["Service A"] -->|"back-pressure applied\nHERE"| B["Service B"] B --> C["Service C"] C --> D["Service D (slow)"] Note["If D is slow, and C simply\nBUFFERS unboundedly instead of\npropagating pressure BACK to B,\nthe problem just moves to C's\nbuffer instead of being solved"]

If Service D (at the end of a chain) is the actual bottleneck, but Service C absorbs the slowdown with its own unbounded internal buffer instead of signaling back to B that it needs to slow down, the back-pressure "stops" at C — B and A keep producing at full speed, and C's buffer just grows instead of D's, delaying (not solving) the exact problem from junior.md, just one hop further back.

True end-to-end back-pressure: every hop propagates the signal¶

sequenceDiagram participant A participant B participant C participant D as D (slow) D->>C: I'm slow, reduce my inbound rate C->>C: applies its OWN back-pressure\nto its upstream C->>B: I'm now also constrained,\nreduce YOUR rate B->>B: applies back-pressure\nto ITS upstream B->>A: reduce your rate too Note over A,D: The slowdown signal propagates\nALL THE WAY BACK to the\nORIGINAL source, A

For back-pressure to genuinely solve (not just relocate) the problem, every intermediate service in the chain must propagate the slowdown signal to its own upstream, rather than absorbing it locally with an ever-growing buffer — this requires every hop in the chain to implement the same credit/pull-based discipline from middle.md, consistently, end to end.

🎯 Senior takeaway: back-pressure is only as effective as its weakest propagating hop — a single service in a multi-hop chain that absorbs pressure locally (buffers unboundedly instead of signaling upstream) breaks the entire chain's end-to-end back-pressure guarantee, even if every other hop implements it correctly. This is analogous to the "one accidentally shared component defeats isolation" lesson from the Deployment Stamps professional page, applied to flow control instead of failure domains.

Test yourself¶

Why does a single non-propagating hop in a multi-hop chain undermine the entire chain's back-pressure guarantee, even if every other hop handles it correctly?
Trace through a 4-hop chain (A→B→C→D) where D becomes slow, and explain what "true end-to-end back-pressure" would look like at each hop.
Why is this problem structurally similar to the retry-amplification discussion from the Retries & Idempotency professional page, even though the direction of the signal (downstream vs. upstream) is opposite?

Continue to professional.md to see how TCP and Reactive Streams formalize back-pressure at the protocol level.