All Posts

I explore the motivation and techniques behind building a vectorized execution engine for distributed queries. Traditional tuple-at-a-time evaluation fails to utilize hardware efficiently at scale. By expressing queries as linear algebraic operations on batches of column vectors using generated kernels, significant performance gains can be achieved through improved data locality, reduced interpretation overhead and better utilization of CPU resources like SIMD units.

I explore approaches to automatically test databases for logic bugs that can severely impact applications and data integrity. Traditional testing struggles with database complexity, so I examine new techniques like Pivoted Query Synthesis, Non-Optimizing Reference Engine Construction and Ternary Query Partitioning. By generating and validating randomized queries, these methods can reveal hidden logic errors not discovered by conventional testing.

This blog post explores the fascinating interplay between Hybrid Logical Clocks (HLC) and Multi-Version Concurrency Control (MVCC) in enabling distributed transactions in CockroachDB. It covers how HLC addresses causality issues and how MVCC handles concurrency conflicts in a distributed setting.

This document analyzes the distributed transaction protocol used in Percolator, a system for large-scale incremental processing deployed by Google. Percolator extends Bigtable to provide multi-row ACID transactions through a standard two-phase commit protocol.

In this post, I analyze several key design aspects of the Google File System presented in the original paper. I discuss the assumptions GFS made regarding workload characteristics and how that informed choices.