Shayon Mukherjee

A database write can stop after memory, a local SSD, an object PUT, or a replicated WAL. Each choice changes latency, durability, and who owns the work left for later.

Building a small filesystem in Rust with metadata in JSON and file contents in plain local files, using FUSE to explore inodes, caching, and what it means for a write to become durable.

GitHub

How Linux demand-pages memory, what mmap does to physical pages, why replacing a mapping breaks direct memory access and shared memory, and how userfaultfd lets you lazily populate memory without destroying the mapping.

A look at what Python

The machines have become powerful enough to write code for us. What does this mean for those of us who love understanding how things work? And how do we keep learning when AI does the heavy lifting?

A simple architecture for BM25 search over object storage using immutable Tantivy shards, stateless indexers and query nodes, and local NVMe caching for sub-second queries.

Physically remove rows in Parquet on S3 with MPU + UploadPartCopy, copy-on-write objects, and a tiny CAS head to make it simple, fast, and compliant.

A thought experiment in building Parquet-like columnar table on S3 with row-level deletes using conditional writes, tombstone files, and a single-object transaction pointer.

Moving large JSON payloads from PostgreSQL TOAST tables to Parquet on S3 with deterministic sharding, row-group pruning, and range-based reads for millisecond point lookups.