~/projects
Systems, tools and storage engines I've built to understand how the real thing works. Everything is open source; most of it is also running in my homelab, which is where the interesting bugs show up.
Fault-tolerant scheduler built on a hand-rolled Raft implementation: leader election, log replication, and at-least-once dispatch with idempotency keys. Cron expressions and DAG dependencies included, because every team eventually reinvents Airflow badly.
1.2M jobs/day on a 3-node cluster, p99 dispatch < 80ms
Bitcask-style storage engine: an in-memory hash index over append-only segment files. Recovers from kill -9 at any write boundary, merges segments in the background, and never mutates a byte in place.
120k ops/s single node, p99 read < 600µs, 8GB dataset
Kafka → ClickHouse pipeline with size- and time-based batching, backpressure propagated end to end, and a dead-letter queue with replay. Rewriting the flush path cut daily part count 4x and made midnight merges boring.
1M events/s on 3 nodes, end-to-end lag < 2s at peak
Prometheus exporter for the parts of PostgreSQL people forget to watch: WAL generation rate, replication slot lag, bloat estimates and lock waits by query fingerprint. One static binary, zero external dependencies.
used by 40+ people on GitHub and one very nervous homelab
Small enough to read in one sitting: desired-state reconciliation loop, gossip-based membership, health-gated rescheduling. Built to understand Kubernetes, not to replace it.
reconciles 50 containers across 5 nodes in < 1s