Azim Afroozeh, Principal Database Engineer at Polars

Azim Afroozeh

Principal Database Engineer @ Polars

about

Principal Database Engineer at Polars. Previously Postdoc at CWI Database Architectures. Lead developer of FastLanes (columnar file format; partially adopted in DuckDB, Parquet, Lance Format) and ALP (lossless float compression; in DuckDB, Parquet, KuzuDB; Best Artifact Award, SIGMOD 2024).

news
Jan 09, 2026 Defended my PhD thesis FastLanes: A Next-Gen File Format (Cum Laude) at Vrije Universiteit Amsterdam.
Jan 05, 2026 Joined Polars as Principal Database Engineer, working on out-of-core execution for workloads exceeding RAM.
Jun 18, 2025 Paper G-ALP accepted at DaMoN 2025.
May 09, 2025 Sven Hielke Hepkema defended his MSc thesis G-ALP: Rethinking GPU Decompression of LightWeight Encodings, co-supervised with Stefan Manegold.
Apr 20, 2025 Started as Postdoctoral Researcher at CWI in the Database Architectures group, continuing development of FastLanes.
Mar 01, 2025 Paper The FastLanes File Format accepted at PVLDB 2025.
Jun 14, 2024 Received Best Artifact Award at SIGMOD 2024 for ALP: Adaptive Lossless Floating-Point Compression.
Jun 01, 2024 Paper Accelerating GPU Data Processing using FastLanes Compression accepted at DaMoN 2024.
Mar 01, 2024 Paper ALP: Adaptive Lossless Floating-Point Compression accepted at SIGMOD 2024.
Jun 01, 2023 Paper The FastLanes Compression Layout: Decoding >100 Billion Integers per Second with Scalar Code accepted at VLDB 2023.
Education
2020 – 2024 PhD in Computer Science (Cum Laude)
Vrije Universiteit Amsterdam & CWI
  • Thesis: FastLanes: A Next-Gen File Format
  • Co-supervised 5 MSc students on FastLanes and ALP:
    • ALP — novel lossless float compression; orders of magnitude faster than prior encodings thanks to data-parallel layouts of FastLanes; published at SIGMOD 2024.
    • Column Correlations — multi-column compression exploiting cross-column correlations; 20% better compression on already heavily compressed files on Public BI benchmark.
    • Nested Encodings — nested-aware compression for FastLanes; 16% better compression on average (up to 90%), matching or beating Zstd-compressed Parquet on the RealNest benchmark.
    • Predicate Pushdown — data-parallel filter evaluation on cascaded encodings; avoids full decompression.
    • G-ALP — extends ALP to GPUs; rethinks how lightweight encodings should work on GPU hardware; ~10× faster than NVIDIA nvCOMP; published at DaMoN 2025.
2018 – 2020 MSc in Computer Science (Cum Laude)
Vrije Universiteit Amsterdam
Experience
2026 – Present Principal Database Engineer — Polars
Out-of-core execution for workloads exceeding RAM.
2025 Postdoctoral Researcher — Database Architectures, CWI
Led development of FastLanes, a SIMD/GPU-friendly columnar file format.
2019 – 2020 Research Intern — Database Architectures, CWI
Built a prototype columnar format as part of MSc thesis; formed the basis for FastLanes.
Projects
2022 – Present FastLanes — SIMD/GPU-friendly columnar file format
40× faster decoding, ~40% better compression than Parquet. Adopted in DuckDB, Parquet, Lance Format; 650+ stars.
2024 ALP — Adaptive Lossless Floating-Point Compression
In DuckDB, Parquet, KuzuDB; proposed for ClickHouse. Best Artifact Award SIGMOD 2024; 150+ stars.
Publications
The FastLanes File Format PVLDB 2025
Azim Afroozeh, Peter Boncz DOI PDF Code
G-ALP: Rethinking GPU Decompression of LightWeight Encodings DaMoN 2025
S. H. Hepkema, Azim Afroozeh, C. Felius, P. Boncz, S. Manegold DOI PDF Code
Accelerating GPU Data Processing using FastLanes Compression DaMoN 2024
Azim Afroozeh, L. Kuffó, P. Boncz DOI PDF Code
Reproducibility Report for ALP SIGMOD 2024
B. N. Bindea, G. Katsogiannis-Meimarakis, L. Zecchini, Azim Afroozeh, L. Kuffó, P. Boncz DOI Code
ALP: Adaptive Lossless Floating-Point Compression SIGMOD 2024 Best Artifact Award
Azim Afroozeh, L. Kuffó, P. Boncz DOI PDF Code
The FastLanes Compression Layout: Decoding >100 Billion Integers per Second with Scalar Code VLDB 2023
Azim Afroozeh, P. Boncz DOI PDF Code
Service
teaching
2022 Teaching Assistant — LSDE, VU Amsterdam