↓ Ir para o conteúdo principal
Scholion
Scholion

← todas as notas

📎 Webclip

Building Distributed Systems in Elixir: Part 9 — Backpressure

The article explains backpressure as the missing control mechanism when a fast Elixir producer outpaces a slower consumer. It uses raw process primitives to show how unbounded mailboxes grow, why that can lead to memory blowups, and how flow control keeps systems from crashing.

Reading notes
#

  • Every process has a private heap and mailbox, and send/2 does not wait for the receiver, so messages can accumulate faster than they are processed.
  • The first experiment shows a slow consumer with Process.sleep(15) and a fast producer filling the mailbox, with Process.info/2 used to inspect queue length and memory.
  • The article links sustained mailbox growth to heap expansion, garbage collection overhead, and eventual OOM termination of the BEAM process.
  • Stop-and-wait flow control uses a correlation reference and an acknowledgment message so the producer sends one item at a time and waits for {:ack, ref} before continuing.
  • This keeps the consumer mailbox at zero or near zero, but it also reduces concurrency and adds round-trip latency.
  • Demand-driven windowing shifts control to the consumer, which asks for a fixed number of items and replenishes demand after each batch.
  • The producer keeps a demand counter, sends batches up to the available demand, and the consumer processes each batch before asking for more.
  • The demand model keeps mailbox size bounded by the window and preserves throughput by pipelining work in batches.
  • The article compares uncoordinated push, stop-and-wait, and demand-driven flow control in terms of control, mailbox risk, throughput, and use case.
  • When upstream sources cannot be throttled, the article says a bounded buffer policy is needed, such as dropping newest, dropping oldest, or rejecting at the boundary.
  • It identifies GenStage, Broadway, and Erlang socket active modes as production equivalents of the demand-driven pattern.
  • The key takeaways are that the BEAM mailbox is unbounded, Process.info/2 can monitor queue length externally, stop-and-wait protects memory but hurts latency, and demand windows keep throughput high while bounding backlog.

Series
#

Building Distributed Systems in Elixir, by krishnadaspc. Research: Distributed Systems Patterns.