Writing · Yahya Ehsan

Field
notes.

software, systems, and the occasional rabbit hole.

Published 02 essays
001 May 30, 2026 15 min read

building a serializable database in rust,
and measuring what it costs

two transactions, two engines, and one of them ends up with nobody on call, a single node engine that detects the anomaly instead of leaving it to you.

two doctors are on call; the rule is that at least one stays on duty. each checks that the other is still on, each signs off, and against most databases you end up with nobody on call. here's the engine that refuses to commit that schedule, built on serializable snapshot isolation, a TLA+ spec checked across 24.6M states, and a deterministic simulator that replays every bug from a seed, and the throughput it pays to do it, measured against RocksDB, LMDB, and sled.

Write skew Two transactions read a shared snapshot and write disjoint keys, forming a read-write cycle. RocksDB commits both and violates the invariant; crackeddb detects the cycle and aborts one, holding the invariant. FIG.01 SHARED SNAPSHOT · t0 alice = on bob = on INVARIANT — AT LEAST ONE ON READ T1 reads alice, bob writes alice = off T2 reads alice, bob writes bob = off rw rw RW CYCLE — SSI ABORTS ONE COMMIT ROCKSDB VIOLATED alice=off, bob=off CRACKEDDB HELD alice=off, bob=on DIAGRAM — WRITE SKEW. TWO TRANSACTIONS READ A SHARED SNAPSHOT AND WRITE DISJOINT KEYS. THE RW CYCLE IS WHAT CRACKEDDB DETECTS AND ROCKSDB DOES NOT.
Read the essay
Rust MVCC / SSI TLA+ Simulation testing Benchmarks
More 01 essay
002

the binding ladder

a rule is only as strong as how hard it is to break. the climb from convention to compile error, and what each rung up costs, measured on deadlock.

May 25, 2026
16 min