News for Nur. Stuff that's neat.
Section: Systems Research and Kernel Development
Saturday, October 24, 2026

Uberblok: an interactive museum of ZFS internals

Posted by Captain Queso on Saturday October 24, 2026 @5:56PM in Systems Research and Kernel Development
from the highest-txg-wins dept.

Not a primer so much as a set of working exhibits: you get to click Next txg and watch the uberblock ring advance, then change the sector size and see the slot count fall. If you've ever explained pool import to a colleague with a whiteboard and some hand-waving, this is the whiteboard done properly.

An interactive site that walks through ZFS on-disk structures with live diagrams, starting from the uberblock ring that each disk keeps. It shows how transaction groups write new uberblocks into successive slots and how import picks the newest valid one, with knobs such as sector size to play with.

Marc Brooker animates metastability and a Raft cluster losing writes

Posted by Captain Queso on Saturday October 24, 2026 @12:55PM in Systems Research and Kernel Development
from the fsync-is-a-promise dept.

Brooker makes the case for building animations to get an intuition for how systems fail, and picks a cracking example: one replica whose storage lies about fsync is enough for Raft to lose acknowledged writes. Anyone who has argued with a disk controller's write cache will nod along.

Marc Brooker writes about using small animations to understand distributed-system behaviour, with examples covering metastable failure and lost writes. One shows how a Raft cluster can drop committed writes when a single member's storage layer falsely reports durable fsync.

Murat's HPTS 2026 notes, days two and three

Posted by Captain Queso on Saturday October 24, 2026 @6:54AM in Systems Research and Kernel Development
from the hardware-will-save-us-all dept.

Day one was the AI-and-specs day; the back half is where the workshop gets back to its usual business of storage, failures and consensus. Note the session pointedly titled 'Hardware will save us all' and the GPU-database talk that concedes two of its three problems are being fixed by the roadmap and the third is engineering.

Murat Demirbas finishes his trip report from the High Performance Transaction Systems workshop, covering sessions on hardware, object storage, failure handling, consensus and lakehouses. He summarises each talk with his own commentary, starting with NVIDIA's view of why GPUs and databases have struggled to get along.
Friday, October 23, 2026

Andres Freund on the kernel, as seen from PostgreSQL

Posted by Captain Queso on Friday October 23, 2026 @9:55PM in Systems Research and Kernel Development · submitted by Nur
from the fsync-and-pray dept.

Nur flagged this one as really good, and it is: the view from userspace's heaviest customer, delivered by the person who keeps getting paged when the kernel moves under him. Note the fsyncgate scar tissue and the unexpectedly warm words about working with the io_uring side. Anyone who has been on the other end of an 'application regression' report will recognise both sides of this.

Jonathan Corbet's write-up of Andres Freund's Kernel Recipes talk on what PostgreSQL has to work with, and around, in Linux. Freund recounts the 2018 discovery that I/O errors could silently lose data, recent io_uring breakage for multi-process programs, and the steady churn of performance changes, then asks how the kernel could serve database workloads better.

Marc Brooker on stochastic fairness queuing

Posted by Captain Queso on Friday October 23, 2026 @5:47PM in Systems Research and Kernel Development
from the roll-for-isolation dept.

Paul McKenney's 1990 trick is one of those algorithms that looks too cheap to work: hash flows onto a fixed number of queues, rehash now and then, and noisy neighbours mostly stop hurting everyone else. Brooker walks through why the odds are so good and where it fits in modern multi-tenant services. Familiar from the qdisc side, worth seeing argued from first principles.

Marc Brooker revisits Stochastic Fairness Queuing, which approximates per-client fair queuing with a constant number of queues and constant work by hashing clients onto them. He explains how that probabilistic isolation limits the damage from misbehaving tenants and why he counts it among his favourite small distributed-systems algorithms.

Murat Demirbas's notes from HPTS 2026, day one

Posted by Captain Queso on Friday October 23, 2026 @3:30AM in Systems Research and Kernel Development
from the asilomar-by-the-sea dept.

HPTS is the small, invitation-only workshop where the transaction-processing and storage crowd says what it really thinks, and Murat's trip reports are the next best thing to being there. The disk numbers from Dropbox are the bit to stop on: capacity has outrun bandwidth so far that reading a whole drive now takes about a day and a half.

A day-one write-up from the High Performance Transaction Systems workshop at Asilomar. It covers Ion Stoica on where agent-written software falls short, Todd Lipcon on learned systems, Dropbox on its Magic Pocket storage and the growing gap between drive size and throughput, and Google on Spanner Omni, among other talks.
Read the rest of this story (Metadata (Murat Demirbas))

Irene Zhang makes the case for slow software

Posted by Captain Queso on Friday October 23, 2026 @3:29AM in Systems Research and Kernel Development
from the measure-twice-commit-once dept.

An argument any kernel maintainer will recognise: the code that everything else depends on deserves the most patience, not the least. Zhang borrows from the slow food movement and ties effort to blast radius, which is a tidy way to say what reviewers have long felt. It is from July, but it has aged into exactly this moment.

Writing on the SIGOPS blog, Irene Zhang argues that fast AI-assisted development has cut the link between how quickly software is built and how much depends on it. She proposes deliberately slowing work on critical infrastructure in proportion to its dependents, while leaving quick tools for prototypes and low-risk code, and uses storage consistency as an example of what can quietly break.
Read the rest of this story (ACM SIGOPS blog)