Engineering Reads

Cognition's essay arguing against multi-agent architectures for coding, paired with a rebuttal on making multi-agent setups actually work.

Tau — an educational open-source project for learning how coding agents are built from scratch.

turbopuffer — disk-based vector + full-text search built on object storage, floated as a comparison point for production RAG stacks.

Anthropic research on a "global workspace" inside Claude — a small subset of internal activity that's consciously accessible, echoing theories of human cognition.

A member's side project: training a small CNN to play Doom with Claude generating most of the code, shared as a learning exercise.

Lynkr — an open-source self-hosted LLM gateway/proxy for coding agents (Claude Code, MCP workflows) with tier routing across providers, semantic caching, and tool-payload compression to cut cost.

"caveman" — a Claude Code skill that strips filler words and hedging from responses to cut token usage by ~65%.

gstack — a packaged set of 23 opinionated Claude Code tools modeled on one founder's personal setup (acting as CEO/designer/eng-manager/QA roles).

TurboQuant — an online vector-quantization algorithm compressing KV cache to 3 bits with no accuracy loss (6x memory, 8x attention speedup), plus Google Research's write-up.

Stripe's "Minions" — homegrown coding agents responsible for 1,000+ merged PRs a week (human-reviewed, agent-written end to end).

A pushback piece on over-delegating architectural decisions to AI: "Claude Is Not Your Architect."

paperdraw.dev — a browser-based system design playground/interview simulator: draw architecture diagrams, simulate traffic and inject chaos.

A web-based System Design Simulator where you drag and drop components and simulate real traffic, failures, latency and scaling.

A friend's AI Agent Orchestrator IDE project (october.dev), pitched as a way to coordinate many agents/tools without typing everything by hand.

Pomodex — a mobile-first AI-powered dev environment: "your computer but on the cloud," controlled from your phone.

OrganizeMyTabs — an AI-powered Chrome extension that auto-groups browser tabs into logical categories.

Stanford/Harvard-attributed paper claim ("Agents of Chaos") that autonomous agents placed in competitive environments drift away from pure performance optimization — shared as a discussion-starter, treat as unverified viral claim.

Foundational reading on Byzantine agreement and probabilistic consensus, shared during a deep-dive phase into distributed computation papers.

Kademlia DHT explained, with a long community back-and-forth on how the "closest subtree" lookup actually converges to the exact node holding a key.

A curated list of classic distributed-systems and CS papers worth reading end to end.

Martin Fowler's catalog of distributed systems patterns, plus the "Fallacies of Distributed Systems" (Deutsch's list).

Amazon's 1998 internal "Distributed Computing Manifesto," published two decades later — the document that pushed Amazon's e-commerce platform toward distributed architecture.

CRDTs explained from first principles, plus a counterpoint on when you probably don't actually need one.

Dolt and OrbitDB — "Git for data": version-controlled, forkable databases exploring peer-to-peer and collaborative data models.

Raft consensus, visualized and re-explained multiple times over the years — from the classic "Secret Lives of Data" visualization to a Project-Loom-based Scala reimplementation and Redpanda's own Raft reconfiguration writeup.

Rendezvous (highest-random-weight) hashing — the lesser-known alternative to consistent hashing used by GitHub's load balancer, Kafka, and others.

A live Cassandra bug report tying back to a "heartbeat handlers must stay non-blocking" lesson from a gossip/phi-accrual failure-detection discussion.

InnoDB internals series and MyRocks (RocksDB as a MySQL storage engine) for anyone wanting to go past the ORM layer.

TigerBeetle — a purpose-built financial transactions database, plus a call for coverage of Viewstamped Replication (the protocol it's built on).

libSQL — a fork of SQLite that's fully open (not just open-source but open-contribution), with a deep engineering write-up covering its driver, protocol and replication.

Redis internals: how expiration/eviction actually works (old sampling method vs. the newer radix-tree approach), keyspace notifications, and the pub/sub wire protocol dissected byte by byte.

Roaring bitmaps explained — the trick behind cutting a 125GB retention-analysis bitmap down to 300MB.

ULID as a sortable, mostly-ordered alternative to UUID for primary keys — with community debate on whether the guarantees actually hold under contention.

How Discord scaled message storage to trillions of rows on Cassandra/ScyllaDB by bucketing channels, and their separate deep-dive into low-latency network disk I/O.

gh-ost and pt-online-schema-change — the two standard "ghost table" approaches for altering huge MySQL tables without taking production down.

Notion's block-based data model, and their separate write-up on sharding the Postgres monolith behind it ("Herding Elephants").

Slack's migration off active-active MySQL onto Vitess for horizontal scaling.

Shopify's approach to balancing MySQL shards with zero-downtime shop migrations.

A deep dive into DynamoDB's actual internals (it diverges quite a bit from the original Dynamo paper) versus HBase's multi-dimensional sorted map model, from a long thread comparing timestamp-based sort keys across the two.

HollandTech's writeup on how Nextdoor evolved its database architecture ("a scaling ladder") as load grew.

Zerodha's long-running case for Postgres over MySQL in production, referenced repeatedly across the years as a "real numbers, real ops experience" example.

DoorDash and Meta's cache-debouncing / stampede-prevention approaches, and a Go "singleflight" library implementing the same coalescing pattern.

Flipkart's memory-tuning writeup for a high-throughput recommendation microservice serving a billion+ recommendations a day.

NATS JetStream's new delayed-message scheduler, and the hashed-wheel-timer data structure it (and many schedulers) are built on.

A CMU-lecture-quality introduction to LSM trees and bloom filters, revisited by different members years apart.

"I/O is no longer the bottleneck" — a 2022 argument that disk latency numbers most engineers still quote are a decade stale, with pushback that at high TPS, I/O still bites.

Latency numbers every programmer should know, refreshed in video form.

KRaft — Kafka's move away from ZooKeeper onto its own Raft-based metadata quorum (3.3 production-ready, 4.0 deprecates ZooKeeper entirely).

How Kafka actually partitions and orders messages — the murmur2-based default partitioner, and why "just add more partitions" isn't free.

Jay Kreps' "The Log" — repeatedly cited as the single best primer on why LinkedIn (and most large systems) end up centered on a commit log.

LinkedIn's Northguard and Xinfra — a from-scratch log storage system built to replace Kafka at LinkedIn's scale, plus a virtualized pub/sub migration layer.

A client-side, key-concurrent Kafka consumer wrapper from the Kafka team itself, useful when strict per-key ordering under high parallelism is the requirement.

A community-maintained resource list specifically for system design prep, plus a 158-page free PDF compiling one prolific poster's LinkedIn content.

A crowd-sourced Google Doc of "important numbers to remember" for back-of-envelope estimation in interviews, plus a GitHub repo of napkin-math techniques and practice problems.

A quick-reference cheatsheet site and Microsoft's Azure cloud design patterns catalog, both bookmarked as go-to references.

Architecture Notes' illustrated deep dives on Redis internals and "things you should know about databases" (indexes, transactions).

A senior-engineer's guide to the system design interview, and a full-text-searchable Google Docs system design textbook maintained by volunteer FAANG engineers.

Cloudflare's DDoS detection/mitigation architecture, and a comic-style walkthrough of how HTTPS actually works.

A deep dive on TCP tunneling / port forwarding, prompted by confusion over how a client "knows" to route through an intermediary server past a firewall.

GitHub's own load balancer, GLB, and its "stretching an IP" trick for active-active failover without a virtual IP.

The AWS us-east-1 service event postmortem, and GitHub's ID-column-overflow outage (2,147,483,647) dissected in video form.

The Facebook/Cloudflare dual perspective on the October 2021 FB global outage (BGP withdrawal cascading into internal tooling lockout).

Spotify's March 2022 global outage, dissected — this time DNS was the fix, not the cause.

How an empty S3 bucket generated a $1,300 bill in two days — a viral cautionary tale about globally-unique bucket names being guessable, plus AWS's public response.

Twitch's two-part series on breaking their monolith into microservices.

PhonePe's Clockwork — the job-scheduling backbone behind 2 billion daily jobs, cited as unusually close to what the community's own scheduler cohort material covers.

Razorpay's account of tracking down a MySQL deadlock worth millions.

Amazon's Builders' Library, referenced repeatedly — particularly the piece on reliability, constant work, and avoiding scale-up during periods of stress.

DiceDB — a real-time reactive database built to solve Redis's lack of client-side change notification, covered in a conference talk.

dkron — a distributed, fault-tolerant job scheduler, recommended repeatedly as the go-to alternative to hand-rolling one on Lambda/K8s CronJobs.

redis-cell — a Redis module implementing GCRA-based rate limiting as a single atomic command, referenced when discussing sorted-set-based rate limiters and their race conditions.

Joel Spolsky's classic "The Absolute Minimum Every Software Developer Must Know About Unicode," re-shared and re-appreciated years apart.

How cron actually works internally, traced back to its original academic data-structure paper, plus one member's own from-scratch implementation write-up.

The Zerodha CTO's essay on chasing new tech stacks without matching business need ("FOMO? YAMO").

A widely-shared explainer of the SOLID principles as illustrated diagrams, for anyone who's read the acronym a hundred times without it sticking.

"Famous laws of software development" — a running list of Conway's Law-style aphorisms worth knowing.