Reading
Things I read and found worth keeping.
November 2026
-
Building a Serverless ClickHouse
Jordi Villar explains how he built ClickHamlet, a serverless ClickHouse that keeps every table in S3 on
plain_rewritabledisks. A C++ router using ClickHouse's own parser sends each query to a single writer or a reader, and a Python controller starts and stops them on demand. He then covers service isolation, performance tuning and hosting on Kubernetes. -
Probabilistic data structures in Go
Anton Zhiyanov explains how probabilistic data structures trade exact answers for less memory and introduces unsure, his Go package implementing them. He covers a cache-line blocked Bloom filter, a count-min sketch with conservative updates, and HyperLogLog using Otmar Ertl's histogram estimator. A merging t-digest and a HeavyKeeper top-k sketch follow, and every structure is benchmarked against other Go libraries.
-
Building an Anycast network for fun
Tyler O'Brien explains how he built a personal IPv6 Anycast network using BGP sessions from Vultr and BuyVM. He covers choosing hosts that offer IP transit, getting an ASN and IPv6 block through a RIPE sponsoring LIR, and adding IRR and RPKI records. He then configures BIRD, measures latency with RIPE Atlas and steers traffic with BGP communities.
-
Inside Replit’s New Sandbox Engine, Part 1: The Path to MicroVMs
Replit explains how it rewrote its sandbox engine, replacing Docker containers on a shared kernel with Cloud Hypervisor microVMs that run runc containers inside. It covers the VMM choice and guest init, then security through a minimal guest kernel, eBPF monitoring and host-side network controls. It ends with remote copy-on-write block storage and warm pools that keep startup fast.
September 2026
-
Go concurrency distilled
Anton Zhiyanov's Go Concurrency Distilled is an interactive mini-book that serves as a quick refresher on concurrent programming in Go, with runnable examples. It moves from goroutines, channels and select to pipeline patterns, timers and context cancellation. Later sections cover data races and the
syncpackage's locks and signals, then finish with atomics and testing concurrent code usingsynctest. -
Anthropic's Hidden Vercel Competitor "Antspace"
AprilNEA's write-up reverse-engineers the Claude Code web runtime using standard Linux tools inside a session. It identifies a Firecracker microVM restored from snapshots under a custom init, then reads an unstripped Go binary whose symbols expose an internal deployment platform called Antspace and an app builder called Baku. It argues Anthropic is building its own AI-native hosting stack.
August 2026
-
How Kubernetes probes work
Sam Rose of ngrok explains Kubernetes startup, readiness and liveness probes using interactive demos on webernetes, his partial TypeScript port of Kubernetes that runs a simulated cluster in the browser. He covers how probes keep traffic off pods that are starting, why liveness checks on shared dependencies cause cascading crash loops, and how probe timing slows Deployment rollouts.
-
PostgreSQL for Everything
Raphael Bauer argues that PostgreSQL can replace many specialized systems, keeping a team's stack simple. He credits its stability and easy setup, then shows full-text search replacing Elasticsearch, JSON with GIN indexes replacing MongoDB and
SKIP LOCKEDtables replacing Kafka queues. Further sections cover Timescale for time series, unlogged tables as a Redis cache and LTREE for hierarchical data.
June 2026
-
Writing a Kubernetes Operator in the Age of AI
Numtide describes how it built a Kubernetes operator for the Multigres distributed Postgres system with AI coding agents. It covers writing the user-facing spec first, turning coding conventions into review skills, and running mechanical work as fresh-context agents whose skills get fixed after every miss. Bug audits get a second verifying agent, while design decisions stay with humans.
-
The feedback loops behind Kubernetes
PlanetScale explains Kubernetes as a framework for feedback controllers, starting from a Postgres cluster run by hand and a Bash watchdog script that forms a closed control loop. It maps that loop to built-in controllers and to operators written with controller-runtime, then covers level-triggered reconciles, informers with work queues, and how a real database operator observes state.
-
Kubernetes Gateway API
Roman Glushko explains the Kubernetes Gateway API as the successor to Ingress, prompted by the retirement of the NGINX Ingress Controller. He traces how Services led to Ingress and lists its limits, such as annotation-based extensions, blurry ownership and unsafe cross-namespace references. He then walks through
GatewayClass, Gateway listeners andListenerSet, and the route types with YAML examples.
May 2026
-
One year of Roto, a compiled scripting language for Rust
Terts Diepraam of NLnet Labs reviews the first year of Roto, a statically typed, JIT-compiled scripting language for Rust programs. Its new language features include loops, enums and a
Listtype, and thelibrary!macro lets Rust types and functions be registered in Rust-like code. It ends with adoption by the Iocaine proxy and plans for hashmaps and tooling. -
How Shamir's Secret Sharing Works
Ente explains Shamir's secret sharing, where a set number of shares recovers a secret and any fewer reveal nothing. The secret sits where a random line crosses the vertical axis, so two points fix it while one fits any value, and higher thresholds use higher degree polynomials. The company uses it as one layer of Legacy Kit recovery.
-
How many sandboxed pods can fit in a Pi?
Nubificus benchmarks plain runc against the sandboxed runtimes gVisor, Kata Containers and urunc on a Raspberry Pi 5 to measure what isolation costs on edge hardware. On single-node K3s with a small C HTTP server, they measure maximum pod density, Kubernetes readiness time and HTTP availability time. They conclude urunc gives VM-level isolation near runc's density and startup speed.
-
Should I Run Plain Docker Compose in Production in 2026?
Philip Miglinci of Distr argues that plain Docker Compose can run production workloads if operators close the gaps it leaves. He covers removing orphan containers, limiting disk use from images and logs, and adding autoheal since health checks never restart anything. Later sections address pinning images by digest, guarding the Docker socket and updating customer hosts through a pull-based agent.
April 2026
-
How I Use Unspent Tokens
The author of Artisincode describes continuous-refactoring, a loop that points a coding agent at a target and keeps each small cleanup only if tests pass. It then covers
taste.mdpreference files built through an agent interview and versioned for upgrades, and a classifier that routes larger jobs to an opt-in migrations workflow with staged plans and human review. -
Migrating from DigitalOcean to Hetzner: From $1,432 to $233/month With Zero Downtime
Isa Yeter recounts moving a production server running MySQL, Nginx sites and GitLab from DigitalOcean to a Hetzner dedicated machine without downtime. The stack was rebuilt on AlmaLinux and files copied with
rsync, whilemydumperand live replication kept MySQL in sync. Shorter DNS TTLs, a reverse proxy on the old host and scripted record changes handled the cutover. -
LLM Wiki
Andrej Karpathy wrote an idea file for LLM agents describing a knowledge base where the model maintains a wiki of linked markdown pages instead of retrieving raw chunks per query. The design separates immutable sources, the LLM-written wiki and a schema file of conventions. Ingest, query and lint workflows keep the wiki current, aided by an index and a log.
-
Building a high-volume metrics pipeline with OpenTelemetry and vmagent
Airbnb engineers describe moving their StatsD metrics onto a Prometheus-based pipeline built around OpenTelemetry. Services dual-wrote StatsD and OTLP through a shared library, using delta temporality for the highest-volume emitters. A sharded vmagent tier of routers and aggregators drops instance labels, and its aggregators flush a zero for each new counter so
rate()stops undercounting sparse counters. -
SSH certificates: the better SSH experience
Jan-Piet Mens explains how an SSH certificate authority, built from one key pair and
ssh-keygen, replaces trust on first use andauthorized_keysfiles. He signs user and host keys, configures the server and the client'sknown_hosts, and limits certificates by principal, command and source address. He then covers revocation, troubleshooting and a small Python service that signs host keys. -
High-Level Rust: Getting 80% of the Benefits with 20% of the Pain
Hamy argues that Rust can be written as a high-level language, keeping its expressive types and speed while avoiding most of its learning curve. He proposes type-first domain modeling, functional logic that clones instead of mutating, and services behind traits shared through
Arc. He then covers the cost of deep clones and when the approach fits. -
Research-Driven Agents: What Happens When Your Agent Reads Before It Codes
SkyPilot adds a literature search phase to the autoresearch loop so coding agents read papers and competing projects before optimizing code. Applied to llama.cpp CPU inference, the agent ran experiments on cloud VMs and shifted from SIMD tweaks to kernel fusions found in other backends and forks. It also covers failed experiments, noisy benchmarks and setup for other projects.