Tombstone
Career: Facebook Real-Time Infrastructure and Bladerunner
2015 โ 2021 ยท age 33โ39
What it was
Six years as a Principal Software Engineer on Facebook's real-time infrastructure, on record through a public SREcon17 talk, a "Request stream" patent filed in 2019, and a SOSP '21 paper on Bladerunner with Jeff as first author. The problem: keep each user's view of a fast-changing social graph fresh, when 80 percent of polls returned nothing. Pylon, a deliberately simple pub/sub, carries only update events, never the data. BRASS runs one small stream processor per application, a few hundred lines of JavaScript on V8, which fetches the payload where privacy is checked and filters, ranks and rate-limits per device. BURST tells every participant about any failure and lets the server rewrite the client's own subscription. Per the paper, over 100 applications ran on it, with 3 billion device subscriptions a minute at peak.
Wins, for the age
- In their thirties, Jeff worked the question the side projects had circled since a chess lobby in 1999 at billion-user scale, and wrote it up at SOSP as first author.
- The team replaced a generic, configurable pub/sub, declared a failure after years, with code: one sandboxed program per application. That arrived alongside Cloudflare Workers, which made JavaScript isolates famous in 2017.
- Promised best effort, reliably signalled, instead of delivery.
- Measured the switch in production: live comment latency at p95 fell from 14 seconds to 6. This ape had finally met a fleet big enough to give real numbers.
What it taught
- The generic version fails and the programmable version works. Every later platform started from "give them a small program", not "give them a knob".
- Privacy checks happen where the data is fetched.
- State the axioms, write down what is slow, and name the builders: first author is coordination credit as much as invention.
Genealogy
Ancestors: HUROX JS: a heartbeat pushing batched updates to a 2007 browser. Play Workspace: one frame protocol spoken by browser, gateway and server. Descendants: Lamancha II: privacy in the type system, started while Bladerunner ran. Adama: one small program per document, fed by an ordered log. Adama Core: failure that degrades to polling, per viewer. Adama Net: rendezvous hashing, brought home.
Epitaph
Here lies a chapter, not a system. Eighty percent of the polls came back empty, so Jeff stopped asking and taught the graph to tell each person what they were looking at, one small program per application.