Choosing QuePaxa, an unusual consensus algorithm
Cloudflare details their choice to adopt QuePaxa, a relatively new consensus algorithm, for their global control plane service Meerkat. The post examines the architectural trade-offs of choosing QuePaxa over established alternatives like Raft, Multi-Paxos, or Viewstamped Replication.
Cloudflare · 14 November 2026 · Read the original →
How We Built TLX: A Hardware-Native GPU Compiler for Production ML at Scale
Meta developed Triton Low-Level Extensions (TLX) to give engineers explicit warp-group and cluster-level orchestration capabilities directly within the Triton compiler. This extension preserves Triton's high-level programming velocity while enabling hardware-native optimizations that yield performance gains in production recommendation workloads.
Meta · 14 November 2026 · Read the original →
Skipper, Airbnb’s embedded workflow engine, is now open source
Airbnb open-sourced Skipper, an embedded workflow engine that runs directly inside service runtimes rather than relying on a standalone orchestration cluster. It allows developers to define durable multi-step processes as standard Java or Kotlin classes with built-in progress persistence and crash recovery.
Airbnb · 14 November 2026 · Read the original →
How Pinterest Improved Analytics Agent Compliance with Loop Engineering
Pinterest engineered loop control constraints for their internal Analytics Agent to reliably diagnose query latency anomalies across data pipelines. The system forces the LLM agent to follow structured diagnostic steps sequentially rather than jumping to conclusions or missing compliance steps.
Pinterest · 14 November 2026 · Read the original →
Intelligent Delivery: Making Every Unit of Compute Serve Better Ads
Meta built Intelligent Delivery to dynamically allocate compute resources across serving stacks, request types, and hardware chips as ML ads ranking models outpace hardware growth. The system evaluates where an extra unit of compute delivers the highest marginal value across the entire request pipeline.
Meta · 14 November 2026 · Read the original →
Extending LLDB to Debug Applications That Span CPUs and GPUs
Meta created a vendor-neutral accelerator plugin architecture inside lldb-server to handle debugging across modern hybrid workloads. The framework represents CPUs and GPUs as distinct sub-processes within a single unified debugging session, separating debugger infrastructure from device-specific execution controls.
Meta · 14 November 2026 · Read the original →
How Zillow runs 100 million daily notifications on AWS Step Functions
Zillow processes over 100 million daily personalized alerts by representing every notification delivery as an independent, short-lived AWS Step Functions execution state machine. This design provides consistent execution behavior, isolated retry logic, and precise campaign tracking across massive daily notification spikes.
AWS Architecture · 14 November 2026 · Read the original →
So much custom infrastructure just to serve ads and send push notifications.
7 stories, every Saturday
Published here every week. Follow by RSS to get it as it lands.