The reference architecture Netflix retired
Where a cross-cutting service rule should live, traced through a decade of Netflix's service-to-service stack, 2015 to 2026, read entirely from its own repositories.
Between 2013 and 2016 the industry copied one company's answer to service-to-service communication: discovery, load balancing, circuit breaking and configuration shipped as client libraries compiled into every service. This guide reconstructs what Netflix did with that stack afterwards, from status notices with commit dates, an abandoned rewrite, closed-unmerged pull requests, operator-filed incident threads, a downstream framework deleting fourteen modules in one release and the parallel decision at gRPC. A reader leaves with a decision rule for where to put a rule that applies everywhere, the measured cost of getting it wrong, and a repeatable method for auditing any vendor's open-source infrastructure before adopting it.
Ribbon's own status notice marks four of its nine published modules as not used internally and says the production ones were wrapped in an internal client, so part of the reference architecture the industry copied was never the architecture the company ran.
What you get out of it
- The components that survived a decade are the ones whose purpose is changing something at runtime: dynamic configuration, and a gateway whose filters are polled from disk and compiled into the running server. Everything that required a redeploy to change a rule was frozen or replaced.
- The stated reason for abandoning the in-process load balancer was multi-language support, and the sidecar built in 2014 to reach non-JVM applications was already the same admission two years earlier.
- The fixed limit lost to the control loop for one reason the source states plainly: under autoscaling a tuned constant goes stale and the service falls over by becoming non-responsive.
- An upstream freeze is not your deadline; the downstream removal is. Ribbon was frozen in April 2016 and deleted from Spring Cloud in January 2021, four years and nine months of running code nobody was maintaining.
- The support commitment is the pull request queue, not the README: Ribbon offers to review community pull requests and has 40 open, the oldest from February 2018, while Hystrix shows 52 open pull requests against 6 open issues.
Scope
Why this, now. The 2013-2016 NetflixOSS diagram is still the microservices reference architecture in circulation, and the repositories now record, with dates, that most of it was frozen, abandoned or replaced by its own authors.
What it does not cover. Netflix's content delivery network, encoding, playback clients, data and machine-learning platforms; the internal systems that replaced the public ones, which are named in status notices but not published; and any engineering blog, conference talk or paper, because the research environment could reach only github.com and raw.githubusercontent.com.
Other field guides
The retry you are counting on was never promised
Reconstructs the sender's side of webhook delivery from GitLab's incident reviews and shipped source, GitHub's staff answers, three open-source deliv…
24 sources · 10 organisations · 4 postmortemsA timeout is not an answer: making retried requests safe
A client that times out on a mutating call holds no evidence about whether it happened, so it must resend, and something has to make the resend harml…
28 sources · 25 organisations · 4 postmortemsYou cannot delete the dual write, only move it
A service commits a change and must announce it, and no mechanism makes both happen or neither. This guide reconstructs where eighteen organisations …
28 sources · 24 organisations · 3 postmortems