Vectorization in Apache Doris: Design, Implementation, Current Status, and Future Plans

Introduction

The article introduces the concept of vectorization, describing how turning single‑value operations into batch operations improves CPU efficiency and has become a major trend in software development, especially for databases.

CPU Vectorization Principles

Modern CPUs support SIMD instructions that can process multiple data elements in a single instruction, e.g., a 128‑bit register can handle four 32‑bit integers simultaneously, providing up to four‑fold speedup compared to scalar execution.

Database Perspective

In a database engine, traditional row‑wise processing handles one tuple at a time, while a vectorized engine processes batches of columnar data, reducing memory accesses and improving cache locality.

Doris Vectorization Design and Implementation

Doris transforms its execution engine from a row‑batch/tuple model to a column‑oriented Block and Column model, redesigns the computation framework to operate on columns, and implements vectorized operators such as aggregation, sorting, and join.

Current Status

Since version 0.15, Doris supports a vectorized engine for single‑wide tables, enabling vectorized scan, sort, agg, and union. Benchmarks show 2‑10× performance gains for typical queries, and the engine can be enabled with set enable_vectorized_engine = true and set batch_size = 4096 .

Future Plans

Key upcoming work includes:

Vectorizing the JOIN operator (currently 30‑40% faster, with further tuning planned).

Vectorizing the storage layer to eliminate row‑wise aggregation and improve memory layout for Date/DateTime, Decimal, and HLL types.

Vectorizing data import to reduce format conversion and leverage SIMD for aggregation during load.

Expanding SIMD support for SQL functions and developing a vectorized UDF framework.

Refactoring basic data types (Date/DateTime, Decimal, String, Array) for better memory usage and SIMD friendliness.

Improving the vectorized execution engine with a cost‑based optimizer to enable deeper inlining and other optimizations.

The team also acknowledges contributions from the ClickHouse community, which provided column‑store and function‑framework code.

Conclusion

The presentation ends with a thank‑you and a call for audience engagement.