I'm a senior backend engineer at Clickpost, working remotely from India. I work on the platform that predicts delivery dates for e-commerce orders: making APIs fast, keeping data consistent, and finding out why something slowed down. Before that I was at OnePlus R&D and TCS, and I did my M.Tech in Computer Science at IIT Bombay.
Outside work I build things to understand how distributed systems actually behave when parts of them fail.
BeeDB is a replicated key-value store in Java 21 with a Raft implementation I wrote by hand. It speaks the memcached protocol, writes every change to a CRC-framed write-ahead log, and only acknowledges a write once a majority has it and it is on disk.
It runs live on three nodes at beedb.subodhlatkar.com. Every few minutes one of them is killed on purpose, and you can watch the other two elect a new leader and the dead one catch up. From the live server:
- about 1,500 writes a second with a p99 of 20 ms, on one 2-vCPU machine
- no acknowledged write lost in the crash tests
- 37 hours of continuous node kills after the latest fix, with the write-ahead log never above 200 KB
The bug behind that last fix was the most interesting one so far: a follower saved the same five entries half a million times, until its log was 478 MB and it could no longer restart.
It started as John Crickett's build your own memcached challenge.
Building BeeDB is a short series for people who have never heard of Raft:
- Why I built a database from scratch
- Following one write through BeeDB
- Writing to disk without lying
- Mistakes that taught me the most
The pinned repos below are mostly from IIT Bombay: a C++ key-value server over gRPC, my M.Tech thesis (BPMN models to Petri nets), a shell for xv6, and a concurrency assignment on sequential consistency.
Java, Python, C++ · Django, Spring Boot, FastAPI · ScyllaDB, PostgreSQL, Redis, Kafka · Raft, write-ahead logs, replication · Docker, Linux, AWS
In Search of an Understandable Consensus Algorithm (Ongaro & Ousterhout) and, next, the Raft dissertation.