Jsoniter Scala

Scala macros that generate codecs for case classes, standard types, and collections to get maximum performance of JSON parsing and serialization.

Latest results of benchmarks on JVMs that compare parsing and serialization performance of jsoniter-scala with AVSystem's scala-commons, Borer, Circe, DSL-JSON, Jackson, Play-JSON, play-json-jsoniter, Spray-JSON, uPickle, weePickle, and zio-json libraries using different JDK and GraalVM versions on the following environment: Intel® Core™ i9-9880H CPU @ 2.3GHz (max 4.8GHz), RAM 16Gb DDR4-2400, macOS Mojave 10.14.6, and latest versions of Amazon Corretto 8/11, OpenJDK 17 (early-access build) *, GraalVM CE 21.1 (dev build) for Java 11/16, GraalVM EE 21.1 (release) for Java 8/11/16.

Latest results of benchmarks on browsers that compares the same libraries on the same environment by the same code which is compiled by Scala.js to ES 5.1 with GCC optimizations applied.

Acknowledgments
Goals
Features and limitations
How to use
Known issues
How to develop

Acknowledgments

This library had started from macros that reused jsoniter (json-iterator) for Java reader and writer but then the library evolved to have its own core of mechanics for parsing and serialization.

The idea to generate codecs by Scala macros and main details were borrowed from Kryo Macros and adapted for the needs of the JSON domain.

Other Scala macros features were peeped in AVSystem Commons and magnolia libraries.

Ideas for the most efficient parsing and serialization of java.time.* values were inspired by DSL-JSON's implementation for java.time.OffsetDateTime.

Other projects that have helped deliver unparalleled safety and performance characteristics for parsing and serialization of floating-point and big numbers:

Schubfach - the most efficient and concise way to serialize doubles and floats to the textual representation
rust-lexical - the most efficient way to parse floats and doubles from the textual representation precisely
big-math - parsing of BigInt and BigDecimal values with the O(n^1.5) complexity instead of O(n^2) using Java's implementations where n is a number of digits

Goals

Safety: validate parsed values safely with the fail-fast approach and clear reporting, provide configurable limits for suboptimal data structures with safe defaults to be resilient for DoS attacks, generate codecs that create instances of a fixed set of classes during parsing to avoid RCE attacks
Correctness: support the latest JSON format (RFC-8259), do not replace illegally encoded characters of string values by placeholder characters, parse numbers with limited binary representation doing half even rounding for too long JSON numbers, serialize floats and doubles to the shortest textual representation without loosing of precision
Speed: do parsing and serialization of JSON directly from UTF-8 bytes to your data structures and back, do it crazily fast without using of run-time reflection, intermediate ASTs, strings or hash maps, with minimum allocations and copying
Productivity: derive codecs recursively for complex types using one line macro, do it in compile-time to minimize the probability of run-time issues, optionally print generated sources as compiler output to be inspected for proving safety and correctness or to be reused as a starting point for the implementation of custom codecs, prohibit serializing of null Scala values and parsing immediately to them in generated codecs
Ergonomics: have preconfigured defaults for the safest and common usage that can be easily altered by compile- and run-time configuration instances, combined with compile-time annotations and implicits, embrace the textual representation of JSON providing a pretty printing option, provide a hex dump in the error message to speed up the view of an error context

The library targets JDK 8+ and GraalVM 19+ (including compilation to native images) without any platform restrictions.

Features and limitations

JSON parsing from Array[Byte], java.io.InputStream or java.nio.ByteBuffer
JSON serialization to Array[Byte], java.io.OutputStream or java.nio.ByteBuffer
Support of parsing from or writing to part of Array[Byte] or java.nio.ByteBuffer by specifying of position and limit
Parsing of streaming JSON values and JSON arrays from java.io.InputStream without the need of holding all input and parsed values in the memory
Only UTF-8 encoding is supported when working with buffered bytes directly but there is a fallback to parse and serialize JSON from/to String (while this is much less efficient)
Parsing of strings with escaped characters for JSON keys and string values
Codecs can be generated for primitives, boxed primitives, enums, tuples, String, BigInt, BigDecimal, Option, Either, java.util.UUID, java.time.* (to/from ISO-8601 representation only), Scala collections, arrays, module classes, literal types, value classes, and case classes with values/fields having any of types listed here
Classes should be defined with a primary constructor that has one list of arguments for all non-transient fields
Non-case Scala classes also supported, but they should have getter accessors for all arguments of a primary constructor
Types that supported as map keys are primitives, boxed primitives, enums, String, BigInt, BigDecimal, java.util.UUID, java.time.*, literal types, and value classes for any of them
Codecs for sorted maps and sets can be customized by implicit Ordering[K] instances for keys that are available at the scope of the make macro call
Core module support reading and writing byte arrays from/to Base16 and Base64 representations (RFC 4648) for using in custom codecs
Parsing of escaped characters is not supported for strings which are mapped to byte arrays, numeric and java.time.* types
Support of first-order and higher-kind types
Support of 2 representations of ADTs with a sealed trait or a Scala class as a base type and non-abstract Scala classes or objects as leaf classes: 1st representation uses discriminator field with string type of value, 2nd one uses string values for objects and a wrapper JSON object with a discriminator key for case class instances
Implicitly resolvable value codecs for JSON values and key codecs for JSON object keys that are mapped to maps allows to inject your custom codecs for adding support of other types or for altering representation in JSON for already supported classes
Type aliases are supported for all types mentioned above
Only acyclic graphs of class instances are supported by generated codecs
Order of instance fields is preserved during serialization for generated codecs
Throws a parsing exception if duplicated keys were detected for a class instance (except maps)
Serialization of null values is prohibited by throwing of NullPointerException errors
Parsing of null values allowed only for optional of collection types (that means the None value or an empty collection accordingly) and for fields which have defined non-null default values
Fields with default values that defined in the constructor are optional, other fields are required (no special annotation required)
Fields with values that are equals to default values, or are empty options/collections/arrays are not serialized to provide a sparse output
Any values that used directly or as part of default values of the constructor parameters should have right implementations of the equals method (it mostly concerns non-case classes or other types that have custom codecs)
Fields can be annotated as transient or just not defined in the constructor to avoid parsing and serializing at all
Field names can be overridden for serialization/parsing by field annotation in the primary constructor of classes
Reading and writing of any arbitrary bytes or raw values are possible by using custom codecs
Parsing exception always reports a hexadecimal offset of Array[Byte], java.io.InputStream or java.nio.ByteBuffer where it occurs, and an optional hex dump affected by error part of an internal byte buffer
Configurable by field annotation ability to read/write numeric fields from/to string values
Both key and value codecs are specialized to work with primitives efficiently without boxing/unboxing
No extra buffering is required when parsing from java.io.InputStream or serializing to java.io.OutputStream
Using black box macros only for codec generation ensures that your types will never be changed
Ability to print all generated code for codecs using a custom scala compiler option: -Xmacro-settings:print-codecs
No dependencies on extra libraries in runtime excluding Scala's scala-library
Releases for different Scala versions: 2.11, 2.12, 2.13
Suppressing of all WartRemover warnings for generated codecs
Support of shading to another package for locking on a particular released version
Patch versions are backward and forward compatible
Support of compilation to a native image by GraalVM
Support of Scala.js 1.0+

There are configurable options that can be set in compile-time:

Ability to read/write numbers from/to string values
Skipping of unexpected fields or throwing of parse exceptions
Skipping of serialization of fields that have empty collection values can be turned off to force serialization of them
Skipping of serialization of fields that have empty optional values can be turned off to force serialization of them
Skipping of serialization of fields which values are matched with defaults that are defined in the primary constructor can be turned off to force serialization of that values
Ability to override names of classes of ADTs and fields using a compile-time annotation
Mapping functions from names of classes and their fields to JSON keys or from names of Java enumeration values to JSON strings and back, including predefined functions which enforce snake_case, kebab-case, camelCase or PascalCase names for all fields in the generated codec
An optional name of the discriminator field for ADTs
Mapping function for values of a discriminator field that is used for distinguishing classes of ADTs
Ability to set a precision, a scale limit, and the max number of significant digits when parsing BigDecimal values
Ability to set the max number of significant digits when parsing BigInt values
Ability to set the max allowed value when parsing bit sets
Ability to set the limit for the number of inserts when parsing sets or maps
Throwing of a compilation error for recursive data structures can be turned off
Throwing of a runtime error when the discriminator is not the first field can be turned off
Ability to parse/serialize Scala enumeration from/to id numbers

List of options that change parsing and serialization in runtime:

Serialization of strings with escaped Unicode characters to be ASCII compatible
Indenting of output and its step
Throwing of stack-less parsing exceptions by default to greatly reduce the impact on performance, while stack traces can be turned on in development for debugging
Turning off hex dumping affected by error part of an internal byte buffer to reduce the impact on performance
Size of the hex dump can be adjusted for bigger or smaller number of 16-byte lines
Preferred size of internal input buffers when parsing from java.io.InputStream or java.nio.DirectByteBuffer
Preferred size of internal output buffers when serializing to java.io.OutputStream or java.nio.DirectByteBuffer
Preferred size of char buffers when parsing string values

For upcoming features and fixes see Commits and Issues page.

How to use

Let's assume that you have the following data structures:

case class Device(id: Int, model: String)

case class User(name: String, devices: Seq[Device])

Add the core library with a "compile" scope and the macros library with "compile-internal" or "provided" scopes to your list of dependencies:

libraryDependencies ++= Seq(
  // Use the %%% operator instead of %% for Scala.js  
  "com.github.plokhotnyuk.jsoniter-scala" %% "jsoniter-scala-core"   % "2.9.1",
  // Use the "provided" scope instead when the "compile-internal" scope is not supported  
  "com.github.plokhotnyuk.jsoniter-scala" %% "jsoniter-scala-macros" % "2.9.1" % "compile-internal"
)

Derive a codec for the top-level type that need to be parsed or serialized:

import com.github.plokhotnyuk.jsoniter_scala.macros._
import com.github.plokhotnyuk.jsoniter_scala.core._

implicit val codec: JsonValueCodec[User] = JsonCodecMaker.make

That's it! You have generated an instance of com.github.plokhotnyuk.jsoniter_scala.core.JsonValueCodec.

Now use it for parsing and serialization:

val user = readFromArray("""{"name":"John","devices":[{"id":1,"model":"HTC One X"}]}""".getBytes("UTF-8"))
val json = writeToArray(User(name = "John", devices = Seq(Device(id = 2, model = "iPhone X"))))

To see generated code for codecs add the following line to your sbt build file:

scalacOptions ++= Seq("-Xmacro-settings:print-codecs")

Full code see in the examples directory.

Use macrolizer to print the code for a selected macro call only.

Also, you can use the following on-line services to generate an initial version of your data structures from JSON samples: json2caseclass, json-to-scala-case-class, and json2classes.

For more use cases, please, check out tests:

Samples for integration with different web frameworks and HTTP servers:

Other usages of jsoniter-scala:

Dijon - support of schema-less JSON using safe and efficient AST representation
play-json-jsoniter - provides the fastest way to convert an instance of play.api.libs.json.JsValue to byte array (or byte buffer, or output stream) and read it back
scalatest-json - Scalatest matchers with appropriate equality and descriptive error messages
tapir - Typed API descRiptions

For all dependent projects it is recommended to use sbt-updates plugin or Scala steward service to keep up with using of the latest releases.

Known issues

There is no validation for the length of JSON representation during parsing.

So if your system is sensitive for that and can accept untrusted input then avoid parsing with java.io.InputStream and check the input length for other ways of parsing.

Scalac 2.11 and earlier versions of Scalac 2.12 and 2.13 have a bug that affects case classes which have 2 fields where the name of one is a prefix for another name that contains a character that should be encoded immediately after the prefix (like o and o-o). You will get compilation or runtime error, depending on the version of the compiler, see details of the issue here.

Use latest versions of Scala 2.12 or 2.13 were the issue was fixed or move a definition of the field with encoded chars (o-o in our case) to be after the field that is affected by the exception (after the o field) like here.

A configuration parameter for the make macro is evaluated in compile-time only and requires no dependency on other code that uses a result of the macro's call, otherwise the following compilation error will be reported:

[error] Cannot evaluate a parameter of the 'make' macro call for type 'full.name.of.YourType'. It should not depend on
        code from the same compilation module where the 'make' macro is called. Use a separated submodule of the project
        to compile all such dependencies before their usage for generation of codecs.

Sometime scalac (or zinc) can fail to compile the make macro call with the same error message for the configuration that has not clear dependencies on other code. For those cases workarounds can be simpler than recommended usage of separated submodule:

use the make macro call without parameters when they match with defaults
use the makeWithoutDiscriminator macro call without parameters when the following configuration is used: CodecMakerConfig.withDiscriminatorFieldName(None)
use the makeWithRequiredCollectionFields macro call without parameters when the following configuration is used: CodecMakerConfig.withRequireCollectionFields(true).withTransientEmpty(false)
isolate the make macro call in the separated object, like in this PR
move jsoniter-scala imports to be local, like here and here
use sbt clean compile stage or sbt clean test stage instead of just sbt clean stage, like in this repo

Scalac can throw the following stack overflow exception on make call for ADTs with objects if the derivation call and the ADT definition are enclosed in the definition of some outer class:

java.lang.StackOverflowError
    ...
	at scala.tools.nsc.transform.ExplicitOuter$OuterPathTransformer.outerPath(ExplicitOuter.scala:267)
	at scala.tools.nsc.transform.ExplicitOuter$OuterPathTransformer.outerPath(ExplicitOuter.scala:267)
	at scala.tools.nsc.transform.ExplicitOuter$OuterPathTransformer.outerPath(ExplicitOuter.scala:267)
	at scala.tools.nsc.transform.ExplicitOuter$OuterPathTransformer.outerPath(ExplicitOuter.scala:267)

Also, internal compiler error can happen during compilation of derived codecs for ADT definitions that are nested in some classes or functions like here

Workaround is the same for both cases: don't enclose ADT definitions into outer classes or functions, use the outer object (not a class) instead.

Scala.js doesn't support Java enums compiled from Java sources, so linking fails with errors like:

[error] Referring to non-existent class com.github.plokhotnyuk.jsoniter_scala.macros.Level
[error]   called from private com.github.plokhotnyuk.jsoniter_scala.macros.JsonCodecMakerSpec.$anonfun$new$24()void
[error]   called from private com.github.plokhotnyuk.jsoniter_scala.macros.JsonCodecMakerSpec.$anonfun$new$1()void
[error]   called from constructor com.github.plokhotnyuk.jsoniter_scala.macros.JsonCodecMakerSpec.<init>()void
[error]   called from static constructor com.github.plokhotnyuk.jsoniter_scala.macros.JsonCodecMakerSpec.<clinit>()void
[error]   called from core module analyzer

The workaround is to split sources for JVM and JS and use Java enum emulation for JS.

Code for JVM:

public enum Level {
    HIGH, LOW;
}

Code for JS:

object Level {
  val HIGH: Level = new Level("HIGH", 0)
  val LOW: Level = new Level("LOW", 1)
  
  val values: Array[Level] = Array(HIGH, LOW)

  def valueOf(name: String): Level =
    if (HIGH.name() == name) HIGH
    else if (LOW.name() == name) LOW
    else throw new IllegalArgumentException(s"Unrecognized Level name: $name")
}

final class Level private (name: String, ordinal: Int) extends Enum[Level](name, ordinal)

Some kinds or versions of browsers can show low performance in runtime when the compiler emits ES 2015 that is a default option for Scala.js 1.0+.

A workaround is using the following configuration for the compiler to produce ES 5.1 output:

scalaJSLinkerConfig ~= { _.withESFeatures(_.withUseECMAScript2015(false)) }

Nested option types like Option[Option[Option[String]]] are not supported for all values. Only None and Some(Some(Some(x: String)))) values can be serialized and then parsed without lost of the info. Some(None) and Some(Some(None)) values will be normalized to None.

A workaround could be using of a custom codec, but it cannot be injected precisely for some specified class field yet.

How to develop

Feel free to ask questions in chat, open issues, or contribute by creating pull requests (fixes and improvements to docs, code, and tests are highly appreciated).

Run tests, check coverage and binary compatibility

sbt clean coverage test coverageReport
sbt clean +test +mimaReportBinaryIssues

BEWARE: jsoniter-scala is included into Scala Community Build for 2.11.x, 2.12.x, and 2.13.x versions of Scala.

Printing of code generated by macros

To see and check code generated by the make macro add the -Dmacro.settings=print-codecs option like here:

sbt -Dmacro.settings=print-codecs clean test

Also, to print code generated by the eval macro use the -Dmacro.settings=print-expr-results option.

Both options can be combined: -Dmacro.settings=print-codecs,print-expr-results

Run JVM benchmarks

Before benchmark running check if your CPU works in performance mode (not a powersave one). On Linux use following commands to print current and set the performance mode:

cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
echo performance | sudo tee /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor

Sbt plugin for JMH tool is used for benchmarking, to see all their features and options please check Sbt-JMH docs and [JMH tool docs](https://openjdk.java.net/projects/code-tools/jmh/

Learn how to write benchmarks in JMH samples and JMH articles posted in Aleksey Shipilёv’s and Nitsan Wakart’s blogs.

List of available options can be printed by:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -h'

Results of benchmark can be stored in different formats: *.csv, *.json, etc. All supported formats can be listed by:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -lrf'

JMH allows running benchmarks with different profilers, to get a list of supported use (can require entering of user password):

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -lprof'

Help for profiler options can be printed by following command (<profiler_name> should be replaced by the name of the supported profiler from the command above):

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -prof <profiler_name>:help'

For parametrized benchmarks the constant value(s) for parameter(s) can be set by -p option:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -p size=1,10,100,1000 ArrayOf.*'

To see throughput with the allocation rate of generated codecs run benchmarks with GC profiler using the following command:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -prof gc -rf json -rff jdk8.json .*Reading.*'

Results that are stored in JSON can be easy plotted in JMH Visualizer by drugging & dropping of your file to the drop zone or using the source parameter with an HTTP link to your file in the URL like here.

On Linux the perf profiler can be used to see CPU event statistics normalized per ops:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -prof perfnorm TwitterAPIReading.jsoniterScala'

Also, it can be run with a specified list of events:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -prof "perfnorm:event=cycles,instructions,ld_blocks_partial.address_alias" TwitterAPIReading.jsoniterScala'

List of available events for the perf profiler can be retrieved by the following command:

perf list

To get a result for some benchmarks with an in-flight recording file from JFR profiler use command like this:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -prof "jfr:dir=target/jfr-reports" -wi 10 -i 60 TwitterAPIReading.jsoniterScala'

You will get the profile in the jsoniter-scala-benchmark/jvm/target/jfr-reports directory.

To run benchmarks with recordings by Async profiler, extract binaries to /opt/async-profiler directory and use command like this:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -prof "async:dir=target/async-reports;output=flamegraph;libPath=/opt/async-profiler/build/libasyncProfiler.so" -wi 10 -i 60 TwitterAPIReading.jsoniterScala'

Now you can open direct and reverse flame graphs in the jsoniter-scala-benchmark/jvmtarget/async-reports directory.

To see list of available events need to start your app or benchmark, and run jps command. I will show list of PIDs and names for currently running Java processes. While your Java process still running launch the Async Profiler with the list option and ID of your process like here:

$ ~/Projects/com/github/jvm-profiling-tools/async-profiler/profiler.sh list 6924
Basic events:
  cpu
  alloc
  lock
  wall
  itimer
Perf events:
  page-faults
  context-switches
  cycles
  instructions
  cache-references
  cache-misses
  branches
  branch-misses
  bus-cycles
  L1-dcache-load-misses
  LLC-load-misses
  dTLB-load-misses
  mem:breakpoint
  trace:tracepoint

Following command can be used to profile and print assembly code of the hottest methods, but it requires a setup of an additional library to make PrintAssembly feature enabled:

sbt 'jsoniter-scala-benchmarkJVM/jmh:run -prof perfasm -wi 10 -i 10 -p size=128 BigIntReading.jsoniterScala'

More info about extras, options, and ability to generate flame graphs see in Sbt-JMH docs

Other benchmarks with results for jsoniter-scala:

comparison with other JSON parsers for Scala mostly on samples from real APIs, but with mapping to simple types only like strings and primitives and results for GraalVM EE Java8 only
comparison with the best binary parsers and serializers for Scala
comparison with different binary and text serializers for Scala
comparison with JSON serializers for Scala on synthetic samples
comparison with JSON parsers for Scala when parsing from/to a string representation
comparison with a state-of-the-art filter that by "building structural indices converts control flow into data flow, thereby largely eliminating inherently unpredictable branches in the program and exploiting the parallelism available in modern processors"

Run Scala.js benchmarks

Use JDK 11+ for building of jsoniter-scala-benchmarkJS module:

sbt -java-home /usr/lib/jvm/openjdk-17 jsoniter-scala-benchmarkJS/fullOptJS

Then open the list of benchmarks in a browser:

cd jsoniter-scala-benchmark/js
open scala-2.13-fullopt.html

The released version of Scala.js benchmarks is available here.

Run compilation time benchmarks

Use the circe-argonaut-compile-times project to compare compilation time of jsoniter-scala for deeply nested data structures with other JSON parsers like argonaut, play-json, and circe in 3 modes: auto, semi-auto, and derivation.

Publish locally

Publish to local Ivy repo:

sbt clean +publishLocal

Publish to local Maven repo:

sbt clean +publishM2

Release

For version numbering use Recommended Versioning Scheme that is used in the Scala ecosystem.

Double-check binary and source compatibility, including behavior, and release using the following command:

sbt release

Do not push changes to GitHub until promoted artifacts for the new version are not available for downloading on Maven Central Repository to avoid binary compatibility check failures in triggered Travis CI builds.

The last step is updating of the tag info in a release list.

pawelkaczor / jsoniter-scala