Four years ago, while doing a code review at work, I was surprised that Flow had not warned about an unnecessary null check. Last month, TypeScript 5.6 released with validation rules that disallows useless nullish and truthy checks which uncovered nearly 100 existing bugs in the top 800 TypeScript repos on GitHub.
The two events are connected because that moment in code review four years ago led me to write the no-constant-binary-expressions
rule which catches a wide variety of bugs. Examples include:
// Expecting empty objects to be falsy
const foo = { ...config } || {};
// Confusing precedence of !
const foo1 = !x == null;
// Confusing ?? or || precedence
const foo2 = x === y ?? true;
The no-constant-binary-expression
rule, in turn, helped inspire the newly added TypeScript validations.
Given the protracted timeline and the many intermediate steps, I thought it would be interesting to reflect on how we got here. How did this observation in one code review snowball into a significant positive impact to developers? And how could the snowball continue to grow?
To answer these questions, it helps to review the timeline of events.
Timeline
- May 2020: I was reviewing a pull request at work that made a nullable value non-nullable. I noticed that the author had left behind an
if
condition that handled the now impossible case where the value wasnull
. I started to wonder why Flow hadn’t automatically pointed that out. - May 2020: I posted to an internal group asking the Flow team about it. The answer was that Flow, like TypeScript, is not sound. For example
arr[x]
is typed as non-nullable but might actually be undefined at runtime. Flow had implemented these checks in the past, but they caused major issues by telling people their null checks were safe to remove when they were not, so they removed the checks. Brad Zacher happened to see the post and chimed in with an observation that a syntactic approach, rather than a type based one, could be safe even if it was less powerful. - August 2020: I implemented the syntactic validation approach as an internal ESLint rule and realized it generalized beyond null checks to all constant comparisons. I ran it in Meta’s mono-repo and found it identified several hundred existing bugs.
- October 2020: Brad Zacher suggested I propose it as a new core rule in ESLint. I did, and they liked the idea.
- November 2021 - April 2022: I rewrote the rule for open source, which took a surprising amount of effort due to different positions on things like JSX and style.
- July 2022: I decided to write a blog post about the new rule. The ESLint team was redoing their site at the time and looking for more blog content, so they asked if I wanted to write the post on the official blog.
- November 2023: The blog post made the front page of Hacker News. The post was likely shared by someone who read a tweet I posted sharing that it was going to be enabled by default in
eslint:recommended
beginning with ESLint v9.0.0. - April 2024: The rule finally ships as enabled by default in ESLint v9.0.0. This change had to wait for a new major version because enabling a new rule in
eslint:recommended
is a breaking change. - July 2024: The TypeScript team, looking to expand upon the idea of checking for a specific type of constant condition, found the blog post and code. They expanded the set of validations they perform to include most of the checks the ESLint rule performs. They found similar results. 94 real errors in the top 800 TS repos. The success of the ESLint rule gave them confidence to enable the check by default in
tsc
. - September 2024: TypeScript 5.6 ships with default validations to disallow constant nullish and truthy checks.
Reflections
There were a number of key factors that contributed to this small observation in code review helping spur a meaningful ecosystem wide improvement:
- Meta’s internal culture empowered me, a random engineer, to have a direct conversation with members of the Flow team. That conversation happened in a venue where Brad Zacher, an ESLint expert, could happen upon it and chime in.
- Meta’s monorepo gave me direct access to a massive codebase which let me easily run early drafts of the rule to assess how impactful the validation would be on a vast amount of real code.
- Meta’s culture of autonomy gave me the freedom to take on a side quest of writing this rule despite it not being part of my team’s responsibility.
- ESLint’s pluggable architecture empowered me to write my own rule and easily deploy it across the whole company without needing to convince any gate keepers.
- The ESLint team’s openness to, and active support of, a new contributor adding a new core rule despite a 2020 policy of only accepting new rules if they relate to new language features.
- Active communication about the work initiated by me, and amplified by the ESLint team, in the form of a blog post and tweets. These allowed the TypeScript team to connect the dots between a more specific request they got about disallowing
if (/regex/)
, and the broader idea of detecting constant conditions. - Not being satisfied with just pointing out a useless null check in code review when I suspected that a fundamental solution to that class of problem was possible. Being somewhat obsessive, I wasn’t satisfied until that solution was enabled not just for me, my team, or my company, but for the broader ecosystem.
What’s Next
It’s taken four years for the ripple of Brad’s initial observation on that internal post to reach this point, but I think the ideas here have the potential to resonate even further:
- TypeScript and Flow could internally track types which they happen to know are sound, and opportunistically report errors based on that data, allowing checks like this to be performed in many more cases.
- Other unsound languages which previously avoided reporting unnecessary checks for the same reason Flow did could use this same approach to catch logic errors.
- More broadly, dead code elimination is a well understood area of compiler design with many codified techniques and approaches. However, they are nearly always applied as optimizations in the compiler backend. I suspect many of these same dead code elimination techniques could be brought to the front-end of the compiler and used to detect and report bugs.
Conclusion
This effort spanned multiple years and multiple organizations.
As a motivated individual within Meta, a for-profit company, I was able to spark a conversation which spawned an idea. I was then able to leverage ESLint to validate that idea internally. Once validated, the ESLint project, a not-for-profit organization, brought the idea to a large audience via it’s broad adoption, recommended rule set, and blog. With the idea documented and validated at scale, engineers at Microsoft, another for-profit company, were able to bring the validations to an even larger audience via their open source project TypeScript.
Each organization played a different, but key, role according to its strengths and position. I believe a common thread throughout this process, which enabled this scale of collaboration, was the active socialization of ideas. From asking questions during code review, to asking questions of local experts, to sharing ideas in public blog posts and tweets, each expanding circle of socialization improved the idea and ultimately helped bring it to the broad audience it has reached today.
Thanks to Brad Zacher for his initial key observation and ongoing encouragement, to Nicholas C. Zakas and Milos Djermanovic for significant contributions to the rule during code review, and to Ryan Cavanaugh for bringing these same types of validation to the TypeScript ecosystem.