Evidence ledger

Field guide: When the internal fork wins: ten years of Netflix retiring the platform libraries it published. Research date 2026-09-21. Every row was fetched in this session.

Corpus limit, stated up front. This session's network policy reached github.com, raw.githubusercontent.com and nothing else. netflixtechblog.com, medium.com, netflix.github.io, usenix.org, arxiv.org, youtube.com and every other host tested returned an egress block, including every host that carries Netflix's engineering writing, its conference talks and its outage reports. There are therefore no engineering blog posts, no talks, no papers and no production incident reports in this ledger, and none in the guide. What remains is the repository record: READMEs, lifecycle metadata files, security advisories, build files, commit history, issue threads and pull requests.

That record is a partial and biased witness, and the bias runs in a specific direction. It is unusually good on what a team told the public about a component's status, and when, and close to silent on what replaced it inside Netflix. Every replacement named in this guide, the gRPC-based RPC stack, the internal Conductor fork, the refactored ConsoleMe, the Kubernetes-based Titus, is described only by the sentence in a README that announces its existence. No Netflix repository describes its design. The guide says so rather than filling the gap.

Two grading notes. First, rows graded postmortem are not outage reports. They are retrospective accounts, written by the team that operated the thing, of why a component stopped: archival announcements that give a stated cause, an end-of-life notice, a security advisory whose remediation section is a decision record, and a discontinuation statement for a rewrite. They are graded at that tier because they carry the same payload as an incident review, the assumption that turned out to be false and what changed afterwards, and the guide labels them as what they are. Second, rows graded adr are files whose purpose is to record a decision (a lifecycle marker, a build file that encodes what survived a deprecation), not prose arguments.

Tiers follow the skill's hierarchy: postmortem, source, adr, casestudy, blog, paper, talk, vendor.

Where the record runs out

Named explicitly, because these gaps are where a reader's risk sits.

• No outage postmortems. Nothing in this corpus describes a production failure at Netflix. Any claim in the guide about operational consequence is an inference from a repository artefact and is labelled as one.
• No description of any replacement. The internal gRPC stack (2016), the internal Conductor fork (2023), the refactored ConsoleMe (2026) and whatever schedules containers after Titus (2022) are each named once, in the notice announcing that the public version is finished, and never described.
• No cost figures. Nothing on this estate carries a spend number, an instance count or a unit cost. The scale figures that exist (Maestro, Metaflow) are self-reported in a README with no date attached to the figure.
• Star and fork counts are adoption proxies, not measurements of use. They are used in the guide only to compare orders of magnitude, and they are dated to the check date.
• falcor is not claimed to be dead. Its lifecycle marker says active and it has 10.6k stars, but the repository has published no releases at all, so the guide reports that pair of facts and draws no conclusion from it.