Sjsonnet

A Scala implementation of the Jsonnet configuration language, running on JVM, Scala Native and JavaScript.

Usage

Sjsonnet can be used from Java and Scala:

<dependency>
    <groupId>com.databricks</groupId>
    <artifactId>sjsonnet_3</artifactId>
</dependency>

// Java
sjsonnet.SjsonnetMain.main0(
    new String[]{"foo.jsonnet"},
    new DefaultParseCache,
    System.in,
    System.out,
    System.err,
    os.package$.MODULE$.pwd(),
    scala.None$.empty()
);

// Scala
sjsonnet.SjsonnetMain.main0(
    Array("foo.jsonnet"),
    new DefaultParseCache,
    System.in,
    System.out,
    System.err,
    os.pwd, // working directory
    None
);

As a standalone executable assembly from the github release page:

$ chmod +x sjsonnet.jar

$ ./sjsonnet.jar
error: Need to pass in a jsonnet file to evaluate
usage: sjsonnet [sjsonnet-options] script-file

  -n, --indent       How much to indent your output JSON
  -J, --jpath        Specify an additional library search dir (left-most wins)
  -o, --output-file  Write to the output file rather than stdout
  ...

$ ./sjsonnet.jar foo.jsonnet

Or from Javascript:

$ node

> require("./sjsonnet.js")

> SjsonnetMain.interpret("local f = function(x) x * x; f(11)", {}, {}, "", (wd, imported) => null)
121

> SjsonnetMain.interpret(
    "local f = import 'foo'; f + 'bar'", // code
    {}, // extVars
    {}, // tlaVars
    "", // initial working directory

    // import callback: receives a base directory and the imported path string,
    // returns a tuple of the resolved file path and file contents or file contents resolve method
    (wd, imported) => [wd + "/" + imported, "local bar = 123; bar + bar"],
    // loader callback: receives the tuple from the import callback and returns the file contents
    ([path, content]) => content
    )
'246bar'

Note that since Javascript does not necessarily have access to the filesystem, you have to provide an explicit import callback that you can use to resolve imports yourself (whether through Node's fs module, or by emulating a filesystem in-memory)

Running deeply recursive Jsonnet programs

The depth of recursion is limited by running environment stack size. You can run Sjsonnet with increased stack size as follows:

# JVM
java -Xss100m -cp sjsonnet.jar sjsonnet.SjsonnetMain foo.jsonnet

# Scala Native
SCALANATIVE_THREAD_STACK_SIZE=100m ./sjsonnet foo.jsonnet

# ScalaJS (Node)
node --stack-size=100m

Architecture

Sjsonnet is implementated as an optimizing interpreter. There are roughly 4 phases:

• sjsonnet.Parser: parses an input String into a sjsonnet.Expr, which is a Syntax Tree representing the Jsonnet document syntax, using the Fastparse parsing library

• sjsonnet.StaticOptimizer is a single AST transform that performs static checking, essential rewriting (e.g. assigning indices in the symbol table for variables) and optimizations. The result is another sjsonnet.Expr per input file that can be stored in the parse cache and reused.

• sjsonnet.Evaluator: recurses over the sjsonnet.Expr produced by the optimizer and converts it into a sjsonnet.Val, a data structure representing the Jsonnet runtime values (basically lazy JSON which can contain function values).

• sjsonnet.Materializer: recurses over the sjsonnet.Val and converts it into an output ujson.Expr: a non-lazy JSON structure without any remaining un-evaluated function values. This can be serialized to a string formatted in a variety of ways

These three phases are encapsulated in the sjsonnet.Interpreter object.

Some notes on the values used in parts of the pipeline:

• sjsonnet.Expr: this represents {...} object literal nodes, a + b binary operation nodes, function(a) {...} definitions and f(a) invocations, etc.. Also keeps track of source-offset information so failures can be correlated with line numbers.

• sjsonnet.Val: essentially the JSON structure (objects, arrays, primitives) but with two modifications. The first is that functions like function(a){...} can still be present in the structure: in Jsonnet you can pass around functions as values and call then later on. The second is that object values & array entries are lazy: e.g. [error 123, 456][1] does not raise an error because the first (erroneous) entry of the array is un-used and thus not evaluated.

• Classes representing literals extend sjsonnet.Val.Literal which in turn extends both, Expr and Val. This allows the evaluator to skip over them instead of having to convert them from one representation to the other.

Performance

Due to pervasive caching, sjsonnet is much faster than google/jsonnet. See this blog post for more details:

• Writing a Faster Jsonnet Compiler

Here's the latest set of benchmarks I've run (as of 18 May 2023) comparing Sjsonnet against google/go-jsonnet and google/jsonnet, measuring the time taken to evaluate an arbitrary config file in the Databricks codebase:

	Sjsonnet 0.4.3	google/go-jsonnet 0.20.0	google/jsonnet 0.20.0
staging/runbot-app.jsonnet (~6.6mb output JSON)	~0.10s	~6.5s	~67s

Sjsonnet was run as a long-lived daemon to keep the JVM warm, while go-jsonnet and google/jsonnet were run as subprocesses, following typical usage patterns. The Sjsonnet command line which is run by all of these is defined in MainBenchmark.mainArgs. You need to change it to point to a suitable input before running a benchmark or the profiler.

Benchmark example:

sbt bench/jmh:run -jvmArgs "-XX:+UseStringDeduplication" sjsonnet.MainBenchmark

Profiler:

sbt bench/run

There's also a benchmark for memory usage:

Execute and print stats:

sbt 'set fork in run := true' 'set javaOptions in run ++= Seq("-Xmx6G", "-XX:+UseG1GC")' 'bench/runMain sjsonnet.MemoryBenchmark'

Execute and pause - this is useful if you want to attach a profiler after the run and deep dive the object utilization.

sbt 'set fork in run := true' 'set javaOptions in run ++= Seq("-Xmx6G", "-XX:+UseG1GC")' 'bench/runMain sjsonnet.MemoryBenchmark --pause'

Laziness

The Jsonnet language is lazy: expressions don't get evaluated unless their value is needed, and thus even erroneous expressions do not cause a failure if un-used. This is represented in the Sjsonnet codebase by sjsonnet.Lazy: a wrapper type that encapsulates an arbitrary computation that returns a sjsonnet.Val.

sjsonnet.Lazy is used in several places, representing where laziness is present in the language:

• Inside sjsonnet.Scope, representing local variable name bindings

• Inside sjsonnet.Val.Arr, representing the contents of array cells

• Inside sjsonnet.Val.Obj, representing the contents of object values

Val extends Lazy so that an already computed value can be treated as lazy without having to wrap it.

Unlike google/jsonnet, Sjsonnet caches the results of lazy computations the first time they are evaluated, avoiding wasteful re-computation when a value is used more than once.

Standard Library

Different from google/jsonnet, Sjsonnet does not implement the Jsonnet standard library std in Jsonnet code. Rather, those functions are implemented as intrinsics directly in the host language (in Std.scala). This allows both better error messages when the input types are wrong, as well as better performance for the more computationally-intense builtin functions, other implementations google/go-jsonnet and jrsonnet implement the Jsonnet standard library in the host language too.

Client-Server

Sjsonnet comes with a built in thin-client and background server, to help mitigate the unfortunate JVM warmup overhead that adds ~1s to every invocation down to 0.2-0.3s. For the simple non-client-server executable, you can use

./mill -i show sjsonnet[3.3.6].jvm.assembly

To create the executable. For the client-server executable, you can use

./mill -i show sjsonnet[3.3.6].server.assembly

By default, the Sjsonnet background server lives in ~/.sjsonnet, and lasts 5 minutes before shutting itself when inactive.

Since the Sjsonnet client still has 0.2-0.3s of overhead, if using Sjsonnet heavily it is still better to include it in your JVM classpath and invoke it programmatically via new Interpreter(...).interpret(...).

Publishing

To publish the JVM version to Maven, make sure the version number in build.sc is correct, then run the following commands:

./mill -i mill.scalalib.PublishModule/publishAll \
    --sonatypeCreds $SONATYPE_USER:$SONATYPE_PASSWORD --publishArtifacts "sjsonnet.jvm[_].__.publishArtifacts" --release true \
    --gpgArgs --passphrase=$GPG_PASSPHRASE,--batch,--yes,-a,-b,--pinentry-mode=loopback