Pedro Holanda

Systems

DuckLake

A data lake and catalog format built on SQL and Parquet. I led it from pre-production 0.1 to 1.0, co-writing the format specification since 0.1 and managing its releases. source

CSV engine

A parallel CSV parser that detects types, delimiters, and dialects on its own. The Pollock benchmark, designed independently of DuckDB, measured it as the most robust parser of any tested system: 9.6 out of 10 configured, 8.4 fully automatic. Our only joint work with the benchmark authors was adding DuckDB to it. source

Arrow & ADBC integration

The zero-copy Arrow bridge between DuckDB and the Python data ecosystem, plus DuckDB's driver for the ADBC standard. It has been the default way to move data in and out of the engine since 2021. source

Async I/O

Read-ahead infrastructure in DuckDB's I/O layer that overlaps fetching with compute, so scans over Parquet and object storage spend less time waiting on the network.

BIGNUM arithmetic

Arbitrary-precision integers for DuckDB, so values past the 64-bit limit stay exact instead of overflowing.

ART index

DuckDB's adaptive radix tree index. I implemented the original index and later its persistent on-disk storage, so it stays durable without losing in-memory lookup speed. source

Python client & UDFs

The foundations of DuckDB's most-used client, and the framework that lets users extend the engine in pure Python. source