JSON With Commas and Comments


Summary: JWCC is a minimal extension to the widely used JSON file format with (1) optional commas after the final element of arrays and objects and (2) C/C++ style comments. These two features make it more suitable for human-editable configuration files, without adding so many features that it’s incompatible with numerous other (deliberate and accidental) existing JSON extensions.

Extensibility

The Peter Principle is the half-joking, half-serious observation that people get promoted to their level of incompetence, because being competent at level N leads to being promoted to level N+1.

My colleague Simon Morris made a similar observation about software complexity:

Software has a Peter Principle. If a piece of code is comprehensible, someone will extend it, so they can apply it to their own problem. If it’s incomprehensible, they’ll write their own code instead. Code tends to be extended to its level of incomprehensibility.

The Many JSON Extensions

There’s a similar story with file formats. If they’re comprehensible, they’ll get extended. JSON (JavaScript Object Notation) is this article’s example. The original specification fits on a single page, either as text or diagrams. The file format is simple and ubiquitous. Therefore, there are many extensions - supersets of JSON. Here’s just a few (including two slightly different extensions both called “JSONC”):

Suprisingly, YAML is also a superset of JSON. Not just conceptually, but also in the sense that valid JSON files are also valid YAML files (although there’s some divergence about whether duplicate keys are legitimate). As a bonus, if you use YAML, then to paraphrase Jamie Zawinski: now you have NO problems.

Wandering Off the Specification

There are also informal supersets-of-JSON in widespread use, sometimes more by accident than by design. The Chromium web browser’s JSON parser goes off-spec in a number of ways. The timeline could have been:

  1. Some developer long ago (perhaps in a yak-shaving hurry) wrote or copy/pasted some parsing code that was accidentally too lenient, allowing a superset-of-JSON. Perhaps they re-used existing code that handled C-style string escapes, like the "\n" in "line\nbreak", without realizing that it also unescaped "\v", valid in a C string but not a JSON string.
  2. People use the software. They write first-party and third-party JSON for it. Some of it is actually malformed (e.g. they have "\v" inside strings) but tests (manual and automatic) usually check that new features work, not that all the slightly-incorrect things are rejected. Nobody notices at the time.
  3. Years pass. Hyrum’s Law slowly kicks in. We can no longer tighten this custom JSON parser implementation to follow the spec more strictly because too many things (in unknown places) will break.

This also affects our ability to replace one JSON library with another. For example, we might want to switch from a C++-based JSON parser to a Rust-based one, because of its security benefits. If the upstream Rust library chooses to follow the spec diligently (which is a perfectly reasonable position) then it would ‘break’ our apps that have inadvertently relied on the previous looser-than-the-spec implementation.

We could carry local patches, but that isn’t free. Upstream fuzz-testing infrastructure only exercises the unmodified library, not our patched flavor. Future upstream changes may also invalidate the downstream patch, possibly in subtle ways. An upstream “this new unsafe block is OK because it’s a private implementation detail and nothing in this crate does X” comment might not be aware that our out-of-tree patch does X to its internals.

Quirks

The Wuffs library approach is to expose quirks: runtime configuration options to go off-spec in various ways so that Wuffs’ implementation can be a drop-in replacement for other implementations, without the need for downstream patches.

Wuffs has 20 JSON quirks so far. As always, there are trade-offs. They’re not free (in terms of maintenance cost) and have super-linear complexity: that file’s comments also has 12 call-outs to the subtleties of combining two particular quirks.

Here’s an example of the emergent complexity when combining two simple-sounding JSON extensions. The first one adds C++-style /* slash-star block comments */ and // double-slash line comments. The second one packs multiple top-level values in a single stream, separated by line breaks.

That second extension - by itself and when holding minified, whitespace-free ‘vanilla’ (non-extended) JSON - plays well with Unix’s traditional line-oriented tools. It is sometimes known as Line-Delimited JSON (LDJSON), Newline-Delimited JSON (NDJSON) and JSON Lines (JSONL). But “one value per line” tools’ assumptions can break if slash-star comments can also contain blank lines.

Here’s another question (let’s call it the ‘end of comment’ question). Is the '\n' at the end of of a // double-slash line comment actually part of the comment? At first, this sounds merely philosophical. Comments are ignored and, in ‘vanilla’ JSON, all whitespace is ignored, so why the distinction?

The ‘right’ answer to that ‘end of comment’ question isn’t obvious, but it can affect whether a line comment at the end of a multi-value stream should end in 1 or 2 '\n' bytes. Ideally the answer should be self-consistent with whether a line comment at the end of file must end with the '\n' or whether the implicit EOF (end-of-file) alone suffices. See also the “Parsing JSON is a Minefield” and “Unintuitive JSON Parsing” articles for how subtle a ‘simple’ format like JSON can be.

Wuffs makes one particular choice for that ‘end of comment’ question. Its particular choice probably isn’t that important, more that it made a concious and documented choice.

Clarity, not Terseness

Some general advice, when designing a new file format or extending an existing one, is keep some room for future extensions. For example, allowing unquoted strings (writing foo instead of "foo"), is certainly convenient, but re-defining undefined or datetime without quotes, from invalid JSON syntax to valid some-extension-of-JSON strings, rules out a future extension adding new ‘keywords’.

CBOR is binary at the wire format level (unlike textual JSON) but naturally extends JSON at the object model level. It also has an undefined concept separate from null, and undefined can be a map key. We couldn’t do the ‘obvious’ CBOR-to-some-extended-JSON conversion if undefined, without quotes, was already repurposed to mean a string.

I find it suprising that, in HOCON, “truefoo parses as the boolean token true followed by the unquoted string foo. However, footrue parses as the unquoted string footrue”.

It can also be helpful for a typo like flase to be picked up early as a syntax error (without needing schemas or type checking) instead of silently accepted (as a string, not a bool). This can otherwise be especially dangerous if further processed in a weakly-typed programming language where any non-empty string is ‘truthy’.

[a b c] is invalid ‘vanilla’ JSON syntax, but in the various extended-JSON variants, is it a list with three 1-byte strings or one 5-byte string? Or is it one 3-byte string because three 1-byte strings are implicitly whitespace-delimited and also then implicitly concatenated? Any particular answer can be consistent in its own world, but different JSON extensions make different choices. This can be confusing when software grows large enough (or gains enough transitive dependencies) to have to speak multiple JSON extensions.

These days, when I’m programming in C/C++ or Go, I often add unnecessary parentheses in expressions like (a * b) + c. Even though they’re redundant because of well-defined operator precedence rules, different programming languages have different precedence rules and getting the precedence wrong can lead to hard-to-spot bugs. The Wuffs language actually rejects a bare a * b + c and you have to parenthesize the multiplication or the addition.

Similarly, for JSON-like documents, I prefer the clarity of either ["a", "b", "c"] or ["a b c"], even if it means a little extra typing. Reading is more important than writing for code and configuration, especially when multiple people or long periods of time are involved.

Introducing JWCC

Having said all of that, here is yet another superset-of-JSON, called JWCC (JSON With Commas and Comments). It is a minimal extension. As its name suggests, there are only two new features:

  • “Commas” lets you optionally have a comma after the final element of an array or an object: [1,2,3,]. When you format one element per line, it’s easier to insert and remove elements (and eyeball the diffs) when you don’t have to fiddle with any commas (or lack of commas) on adjacent but otherwise unrelated lines.
  • “Comments” lets you have C++-style /* slash-star block comments */ and // double-slash line comments, anywhere where ‘vanilla’ JSON allows whitespace. Line comments must end with a '\n' byte, even at the end of the file.

To be clear, while every JSON file is valid JWCC, this is a new file format. It just happens to be very familiar if you (or your software) already speak JSON. Yes, Doug Crockford deliberately removed comments from JSON but people keep putting them back in. If we’re going to have comment-enriched JSON (e.g. for human-editable configuration files), we might as well have a standard one. Cue XKCD #927 “Standards”.

C/C++ Implementation

Wuffs’ JSON library (availble as a C or C++ API) can decode either ‘vanilla’ JSON or JWCC, using its quirks mechanism. jsonptr is a command line tool (a JSON formatter) that uses this library. By default, it speaks spec-compliant ‘vanilla’ JSON:

$ echo '[1,2,/*hello*/3,]' | jsonptr
[ 1, 2
json: bad input

It has a JWCC mode:

$ echo '[1,2,/*hello*/3,]' | jsonptr -jwcc
[ 1, 2, /*hello*/ 3,
]

It can also convert from JWCC syntax to ‘vanilla’ JSON syntax, for piping into other tools that only speak the latter:

$ echo '[1,2,/*hello*/3,]' | jsonptr -input-jwcc
[ 1, 2, 3
]

Go Implementation

In a case of parallel evolution, Tailscale already have a Go implementation of this format. They call it HuJSON - Human JSON.

Published: 2021-02-22