Hakana: Taking Hack Severely – Slack Engineering

TL; DR: We’re saying a brand new open supply sort checker for Hack, known as Hakana.

Slack launched in 2014, constructed with a number of love and in addition a number of PHP code.

We began migrating to a special language known as Hack in 2016. Hack was created by Fb after they’d struggled to scale their operations with PHP. It provided extra type-safety than PHP, and it got here with an interpreter (known as HHVM) that might run PHP code quicker than PHP’s personal interpreter.

A lot has modified in PHP-land since we switched. PHP is quicker than it was once, and it has borrowed quite a few Hack options (akin to constructor property promotion). Quite a lot of the PHP group has additionally embraced sort checking — there are actually some nice third-party sort checkers to select from.

Sticking with Hack has given us entry to further runtime pace boosts, performance-enhancing language constructs like `async`, and a typechecker that’s extra strict by default than PHP typecheckers. However we’ve missed out on options offered by PHP typecheckers, together with the flexibility to customise sort inference guidelines to seek out points particular to our codebase and automatic safety vulnerability detection.

Slack has a whole bunch of builders writing Hack. We need to give them the absolute best expertise, so final 12 months we began constructing a sort checker that might fill these gaps.

We’ve dubbed that static evaluation device Hakana, and it’s now available on GitHub!

Hakana relies on Psalm, an open-source PHP static evaluation device I created, and it’s written in Rust. Hakana re-uses a Hack parser that’s bundled with the Hack interpreter.

A bonus of writing it in Rust: with a little bit of prodding, Hakana can run nearly wherever. For instance, it runs in your web browser by way of WASM.

How we use Hakana

At Slack we run Hakana in CI to implement good code conduct in a variety of areas. Right here’s an incomplete record:

• It prevents unused features and unused non-public strategies.
• It prevents unused assignments inside closures.
• It detects each unattainable and redundant type-checks.
• It warns us about potential SQL-injection assaults and cross-site scripting vulnerabilities (extra on this beneath).
• It prevents misuse of inner Slack APIs (by way of plugin hooks).

We additionally use Hakana to automate type-aware API migrations (once more by way of plugin hooks) and to delete unused features in bulk. Because of Rust, these whole-codebase migrations are comparatively fast.

Safety

PHP makes it very easy to make a dynamically-rendered web site. PHP additionally makes it very easy to create an totally insecure dynamically-rendered web site.

Hack improves on this barely, by supporting a system for producing HTML output known as XHP. XHP is secure-by-default towards cross-site scripting assaults, nevertheless it doesn’t cease you from leaking buyer information, and Hack doesn’t forestall you from taking pictures your self within the foot with a variety of different safety vulnerabilities.

For a number of causes (together with compliance obligations) Slack wanted a device that might uncover these vulnerabilities. Psalm, the kind checker that Hakana relies on, already does safety evaluation, so it was comparatively easy so as to add safety evaluation to Hakana as effectively.

Hakana isn’t the primary safety evaluation device for Hack — for years, Fb has been utilizing an inner, closed-source device called Zoncolan — however Hakana is the primary that everybody can use.

Hakana works in a lot the identical method as Zoncolan. It examines how information can move between completely different features in a codebase, and checks if attacker-controlled information can present up in locations it shouldn’t.

Up to now, Hakana has discovered quite a few exploitable vulnerabilities in manufacturing code at Slack (that had been instantly mounted, and we checked our logs to make sure that the vulnerabilities had not really ever been exploited).

Safety within the sort system

Hakana’s safety evaluation mode is a type of interprocedural evaluation — it seems on the method information flows between features. Hakana additionally helps detecting one sort of vulnerability (SQL injection) by way of intraprocedural evaluation, simply by inspecting varieties at operate boundaries.

To do that, Hakana borrows the idea of literal string varieties from Psalm. In Hack code we will outline a sort alias:

<<HakanaSpecialTypesLiteralString()>>
sort db_query_string = string;

Although the official Hack typechecker simply treats this as a string, Hakana treats it as a particular sort `literal-string`, a subtype of string that may solely be concatenated or interpolated with different literal-strings. Passing a string right into a operate that expects a literal-string causes Hakana to emit an error:

operate get_id_query(string $id): db_query_string 
    return "choose * from customers the place id = '$id'";
    // Error: The kind `string` is extra basic
    // than the declared return sort `literal-string`

Extending Hakana

PHP static evaluation instruments are typically highly-customisable. Customisable static evaluation is basically helpful for PHP, as a result of its interpreter permits plenty of tips (e.g. magic methods) that may confound one-size-fits-all static evaluation instruments.

Most of these tips don’t work in Hack code, so there’s far much less of a necessity for customisable static evaluation. Even so, we’ve discovered a number of worth in extending Hakana with plugins.

For instance, we use a customized plugin to inform Hakana {that a} methodology name on our inner End result object, $some_result->is_ok(), is equal to the extra verbose $some_result is ResultSuccess<_> typecheck.

We additionally use customized plugins to carry out type-aware migrations at pace throughout your complete codebase.

Constructing a plugin system for Rust isn’t easy — our customized model of Hakana wraps the open-source core as a library, utilizing its plugin hooks the place obligatory — nevertheless it’s an vital function for us.

Velocity

Hakana was tailored from Psalm, a sort checker written in PHP. Whereas PHP is quick for an interpreted language, it might’t compete with a compiled device. To research a codebase of Slack’s measurement — many hundreds of thousands of traces of code — we would have liked a device that’s as quick as potential.

Hakana, written in Rust, runs about 5x quicker than the PHP-based device on which it’s modeled.

We haven’t spent a lot time tuning efficiency, however when analyzing the complete Slack codebase (about 5 million traces of code) efficiency is on par with the official Hack typechecker. That’s ok for us, for now.

Why open-source Hakana?

Hack type of seems like PHP however with extra varieties. In that respect it’s been in comparison with TypeScript. However in contrast to TypeScript, which compiles all the way down to JavaScript, Hack requires its customers to vary a number of their server infrastructure too.

As a result of that top switching value, at present Hack is just used at a couple of corporations. It’s potential that no one else except for Slack may have a purpose to make use of Hakana.

Even so, there are a couple of the reason why we predict it’s value open-sourcing:

• 🔍 The broader programming language group might have helpful enter — particularly in terms of safety evaluation.
• 🤝 Psalm, one other open-source device, was the premise for Hakana. By open-sourcing our personal device we’re returning the favor.
• 🏢 Although it wouldn’t be straightforward, corporations with extraordinarily massive PHP codebases would possibly contemplate forking Hakana and altering it to investigate PHP code.

Conclusion

Slack is an indispensable device for hundreds of thousands around the globe. Hack is now an integral a part of Slack, and so we’ve created a device that we hope will grow to be indispensable to our Hack builders. We’re open-sourcing it today with the hope that others will discover it helpful and fascinating.