A library, on shelves.

The paper that introduced logical clocks. Required reading; everything later refers back.

The threat model behind every blockchain consensus algorithm — and a clean introduction to fault tolerance.

The shorter, clearer version. Read this before the original Paxos paper.

Designed to be easier to teach than Paxos. Powers etcd, Consul, CockroachDB.

Consistent hashing, vector clocks, eventual consistency — the eventually-consistent KV pattern in one paper.

The blueprint for HBase, Cassandra, and a generation of wide-column stores.

TrueTime + Paxos at planetary scale. Foundational for CockroachDB, Yugabyte, AlloyDB.

Strongly eventual consistency without coordination. The math behind every collaborative editor.

Proves no deterministic protocol can guarantee consensus in an asynchronous network with even one faulty node. The boundary every consensus algorithm has to negotiate.

The pre-CAP framing. Introduces "harvest" (fraction of data returned) and "yield" (queries answered) as orthogonal availability dimensions.

Why nothing is exactly-once and how to design around it. Required reading for anyone doing distributed messaging or workflows.

The paper that founded information theory. Channel capacity, entropy, the bit. Foundational for everything.

The 1986 internet collapse and the slow-start + AIMD response that fixed it.

Replaces packet-loss as the congestion signal with bandwidth + RTT modelling. Now default in Google's edges and Linux kernels.

The HTTP/3 substrate. Multipath, 0-RTT, integrated TLS, connection migration.

Onion routing, circuit construction, hidden services. The model for every modern privacy network.

The architectural principle that shaped TCP/IP: put functionality at the endpoints, not in the network. Required reading for any protocol design.

Google's consistent-hash L4 load balancer. The Maglev hashing construction is the reference everyone copies.

The successor to BBR — adds explicit congestion notification handling and better fairness with loss-based flows.

The Transformer paper. Every modern LLM is a descendant.

ResNet. The skip-connection idea that lets neural nets go past a few dozen layers.

The pre-train then fine-tune paradigm that defined NLP for half a decade.

The paper that made "in-context learning" a central idea — and made scale a deliberate strategy.

Bagging + random feature subsets. The most-deployed tabular ML method, ever.

Two networks, one chasing the other. The mathematics of synthesis.

12 hard-won lessons every practitioner should know. Pithy, opinionated, and right.

The default optimiser used in 80%+ of deep learning papers since 2015. Adaptive moments per parameter.

Train a small student to match a large teacher's soft outputs. Foundation of every modern compressed-LLM technique.

The CSP paper. The conceptual foundation for Go channels, Erlang processes, and most modern actor systems.

The Hindley-Milner type system. Why ML, Haskell, OCaml, and Rust can infer types so well.

Treat each value as used exactly once. The intellectual ancestor of Rust's ownership system and Haskell's linear types.

A C compiler proven correct in Coq. Showed that mechanically verified production-quality compilers are possible.

Multi-volume Coq-based introduction to programming languages, type theory, and formal verification.