Types for Tables: A Language Design Benchmark

TL;DR

This paper introduces a standardized benchmark for table programming languages, focusing on expressive power and diagnostic quality of rich types, to improve comparability and evaluation of different systems.

Contribution

It creates a standard library of table operations with detailed type specifications and a suite of erroneous programs for benchmarking expressiveness and error feedback quality.

Findings

Established a common benchmark for table language expressiveness.

Developed a suite of erroneous programs to evaluate error diagnostics.

Proposed a datasheet format for systematic comparison of implementations.

Abstract

Context: Tables are ubiquitous formats for data. Therefore, techniques for writing correct programs over tables, and debugging incorrect ones, are vital. Our specific focus in this paper is on rich types that articulate the properties of tabular operations. We wish to study both their expressive power and _diagnostic quality_. Inquiry: There is no "standard library" of table operations. As a result, every paper (and project) is free to use its own (sub)set of operations. This makes artifacts very difficult to compare, and it can be hard to tell whether omitted operations were left out by oversight or because they cannot actually be expressed. Furthermore, virtually no papers discuss the quality of type error feedback. Approach: We combed through several existing languages and libraries to create a "standard library" of table operations. Each entry is accompanied by a detailed specification of its "type," expressed independent of (and hence not constrained by) any type language. We also studied and categorized a corpus of (student) program edits that resulted in table-related errors. We used this to generate a suite of erroneous programs. Finally, we adapted the concept of a datasheet to facilitate comparisons of different implementations. Knowledge: Our benchmark creates a common ground to frame work in this area. Language designers who claim to support typed programming over tables have a clear suite against which to demonstrate their system's expressive power. Our family of errors also gives them a chance to demonstrate the quality of feedback. Researchers who improve one aspect -- especially error reporting -- without changing the other can demonstrate their improvement, as can those who engage in trade-offs between the two. The net result should be much better science in both expressiveness and diagnostics. We also introduce a datasheet format for presenting this knowledge in a methodical way. Grounding: We have generated our benchmark from real languages, libraries, and programs, as well as personal experience conducting and teaching data science. We have drawn on experience in engineering and, more recently, in data science to generate the datasheet. Importance: Claims about type support for tabular programming are hard to evaluate. However, tabular programming is ubiquitous, and the expressive power of type systems keeps growing. Our benchmark and datasheet can help lead to more orderly science. It also benefits programmers trying to choose a language.