Own the layer above the engine: ten years of LinkedIn's data platform

Reconstructs a decade of data-platform decisions at LinkedIn from artefacts rather than announcements: archive notices, dated release listings on Maven Central, PyPI, npm, Docker Hub and the Go proxy, published design proposals, an abandoned pull request and five security advisories. A reader leaves able to date any organisation's platform decisions from its registries, to argue the precompute, query and embed choices on a read path with published latency and hop counts attached, and to run a dependency audit that catches the projects whose maintenance ended without a notice.

What you get out of it

• The software LinkedIn still writes is not engines but the layer directly above them: SQL dialect translation, portable user-defined functions, a table control plane over Iceberg, and a planner that compiles a materialized view into pipelines across Kafka, Flink and Venice.
• The read path is a published price list rather than a single design: two hops under 10ms, one hop under 2ms, zero hops under 1ms with the dataset resident in the process, and the stated reason a hop is expensive is the garbage-collected runtime rather than the network.
• Owning the layer above an engine does not keep you out of the engine: LinkedIn maintains a Kafka fork and an Iceberg fork, and on 28 August 2026 it published releases on two parallel Iceberg lines on the same day.
• Absence of an archive notice is not evidence of maintenance. Only four repositories in the organisation are archived, while Databus carries no notice, Feathr's README still claims production use three years after its last artefact, and a LinkedIn container image with 123.7 million pulls stopped being updated in April 2024.
• With no published postmortems, the incident record is the advisory database, where the same trusted-network default keeps reappearing: script execution on by default in 2022, a path-shaped authentication bypass in 2025, and an unauthenticated agent tool server in front of the same engine in 2026.

Scope

Why this, now. Rest.li, the framework that carried LinkedIn's service layer since 2012, went read-only on 22 July 2026, and its final release three weeks earlier was work to make it speak the protocol of its replacement.

What it does not cover. Espresso and the primary-data stores, the feed, search and advertising serving stacks, model training, and cost, because no public artefact in this corpus carries a cost figure.