Implementing Window Functions in a Column-Store with Late Materialization (Extended Version)

TL;DR

This paper presents an implementation of window functions in a disk-based column-store using late materialization, reducing memory usage and applying segment trees for RANGE-based functions.

Contribution

It introduces a novel approach combining late materialization with window functions and segment trees in a column-store database.

Findings

Significantly lowers memory footprint for window function evaluation.

Efficiently computes RANGE-based window functions using segment trees.

Demonstrates practical benefits in a disk-based column-store environment.

Abstract

A window function is a generalization of the aggregation operation. Unlike aggregation, the cardinality of its output is always the same as the cardinality of input. That is, the semantics of this operator imply computing values for extra attributes for each row, depending on its context, either expressed by a sliding window or a previously evaluated row. Window functions are a very powerful tool, which is also popular among data analysts and supported by the majority of industrial DBMSes. It allows to gracefully express quite complex use-cases, such as running sums and averages, local maximum and minimum, and different types of ranking. Since they can be expressed without self-joins and correlated subqueries, their evaluation can be performed much more efficiently. In this paper we discuss an implementation of window functions inside a disk-based column-store with late materialization. Late materialization is a technique that aims to keep tuple reconstruction back from individual columns as long as possible. Initially popular in the late 00's, it is rarely considered nowadays. However, in case of window functions it allows to substantially lower memory footprint. Another contribution of this paper is the application of a segment tree to computing RANGE-based window functions.