Treat backpressure as a contract between producer and consumer, not a nice-to-have. Prefer pipe or pipeline so pause/resume and destroy propagate; if you wire data and drain by hand, always honor the boolean from write and wait for drain before the next chunk. Bound highWaterMark for objectMode streams so a slow sink cannot balloon RSS. Destroy both ends on error, clear listeners, and avoid buffering an entire file before any write — that pattern is what turns a large upload into a leak under load.

Ops prove-out on a canary: stream a multi-gigabyte payload through the pipeline while watching RSS and event-loop lag. RSS should plateau near the configured buffer budget, not climb with file size. Throttle the consumer artificially and confirm the producer pauses; remove the throttle and confirm drain resumes without duplicate chunks. Abort mid-stream and verify both sockets and file descriptors close with no leftover listeners on the next request.