From Code to Cables: Rethinking Reliability in Modern Systems
Talk 1: Merging at Scale: From Broken Builds to Green Mainline
In monorepos, “green” pull requests don’t always guarantee a green mainline. At Uber’s scale, with thousands of engineers and daily commits, independent changes pass builds but fail after merging, leading to integration failures, lost productivity, and slower development. This talk explains how Uber addressed this by rethinking the merge process. We introduce SubmitQueue, a predictive merge system that validates future mainline states before changes merge. It combines speculative execution, building only the components affected by a change, and probabilistic prioritization to focus builds on changes most likely to succeed, keeping the mainline green. We’ll cover:
Why passing builds don’t guarantee safe merges
How speculative merge queues detect integration failures early
Techniques to scale merging (target-based builds, pruning, probabilistic speculation)
Optimizations like bypassing large diffs to reduce wait times
Lessons from operating at Uber scale
This session shares practical approaches to keeping the mainline green while maintaining development velocity.
About Speakers:
Preetam Dwivedi is a Staff Engineer on Uber’s Developer Platform team, leading systems for code hosting, review, merge queues, and artifacts. He specializes in distributed systems and scaling developer infrastructure, with a focus on modernizing workflows and cloud migrations.
Manjari Akella is a Senior Software Engineer on the same team, focused on improving developer productivity. She has contributed to Uber’s Go monorepo, enhanced Go libraries, and worked on artifact management and merge queues.
Talk 2: The Internet is truly a series of tubes
In 2006, Senator Ted Stevens was ridiculed for describing the internet as "a series of tubes." He was more right than anyone gave him credit for. In this talk, we'll peel back the abstractions and follow packets through the real, physical infrastructure underneath: submarine cables on the ocean floor, fiber optics thinner than a human hair, internet exchange points you might walk by every day, and data centers built near the Arctic Circle because your containers generate real heat that someone has to deal with.
We'll look at why the speed of light sets a hard floor on your latency that no code optimization can beat, how network operators cram terabits of data through a single strand of glass, and what you're really choosing when you pick a cloud region. We'll run a traceroute and watch a packet hop across continents in real time and you'll see exactly where the latency comes from.
You'll find out why a wireless API call still travels thousands of kilometers through glass, why the geography of cables creates political chokepoints, and why your five-nines SLA is one inattentive backhoe operator away from being meaningless.
The cloud lets us stop thinking about all of this, and that's mostly a good thing. But understanding the physical layer underneath changes how we reason about latency, redundancy, and failure modes.
About Speaker:
Amaar Zuberi is a Distinguished Engineer at GEICO, specializing in network and systems engineering, with a focus on identity-first security and cloud efficiency. He has previously worked at Cloudflare, Intel Corporation, and Verizon Communications.
Dont forget to RSVP and join us at the event location to learn more about the topics in depth along with your free pizza!.
Thanks Uber for sponsoring the event.