circe is a JSON library for Scala. The rest of this page tries to give some justification for its existence. There are also API docs.
circe's working title was jfc, which stood for "JSON for cats". The name was changed for a number of reasons.
circe is published to Maven Central and cross-built for Scala 2.10 and 2.11, so you can just add the following to your build:
libraryDependencies ++= Seq(
"io.circe" %% "circe-core" % "0.1.1",
"io.circe" %% "circe-generic" % "0.1.1",
"io.circe" %% "circe-jawn" % "0.1.1"
)
Then type sbt console
to start a REPL and then paste the following (this will also work from the
root directory of this repository):
scala> import io.circe._, io.circe.generic.auto._, io.circe.jawn._, io.circe.syntax._
import io.circe._
import io.circe.auto._
import io.circe.jawn._
import io.circe.syntax._
scala> sealed trait Foo
defined trait Foo
scala> case class Bar(xs: List[String]) extends Foo
defined class Bar
scala> case class Qux(i: Int, d: Option[Double]) extends Foo
defined class Qux
scala> val foo: Foo = Qux(13, Some(14.0))
foo: Foo = Qux(13,Some(14.0))
scala> foo.asJson.noSpaces
res0: String = {"Qux":{"d":14.0,"i":13}}
scala> decode[Foo](foo.asJson.spaces4)
res1: cats.data.Xor[io.circe.Error,Foo] = Right(Qux(13,Some(14.0)))
No boilerplate, no runtime reflection.
Argonaut is a great library. It's by far the best JSON library for Scala, and the best JSON library on the JVM. If you're doing anything with JSON in Scala, you should be using Argonaut.
circe is a fork of Argonaut with a few important differences.
circe depends on cats instead of Scalaz, and the core
project has only two
dependencies: cats-core and export-hook (a lightweight mechanism for cleaner generic
type class instance derivation).
Other subprojects bring in dependencies on Jawn (for parsing in the jawn
subproject), Shapeless (for automatic codec derivation in generic
),
and Twitter Util (for tools for asynchronous parsing in async
), but it would be possible
to replace the functionality provided by these subprojects with alternative implementations that use
other libraries.
circe doesn't include a JSON parser in the core
project, which is focused on the JSON AST, zippers,
and codecs. The jawn
subproject provides support for parsing JSON via a Jawn
facade. Jawn is fast, it offers asynchronous parsing, and best of all it lets us drop a lot of the
fussiest code in Argonaut.
circe doesn't use or provide lenses in the core
project (or at all, for now). This is related to
the first point above, since Monocle has a Scalaz dependency, but we also feel that it
simplifies the API. We'd consider adding lenses in a subproject if Monocle (or something similar)
gets ported to cats.
circe does not use macros or provide any kind of automatic derivation in the core
project. Instead
of Argonaut's limited macro-based derivation (which does not support sealed trait hierarchies, for
example), circe includes a subproject (generic
) that provides generic codec derivation using
Shapeless.
This subproject is currently a simplified port of argonaut-shapeless that provides fully automatic derivation of instances for case classes and sealed trait hierarchies. It also includes derivation of "incomplete" case class instances (see my recent blog post for details).
circe aims to simplify Argonaut's API by removing all operator aliases. This is largely a matter of personal taste, and may change in the future.
The Argonaut documentation is good, but it could be better: to take just one example, it can be hard
to tell at a glance why there are three different Cursor
, HCursor
, and ACursor
types. In this
particular case, circe introduces an abstraction over cursors that makes the relationship clearer and
allows these three types to share API documentation.
I'd like to provide more complete test coverage (in part via Discipline), but it's early days for this.
circe aims to be more focused on performance. I'm still experimenting with the right balance, but I'm open to using mutability, inheritance, and all kinds of other horrible things under the hood if they make circe faster (the public API does not and will never expose any of this, though).
My initial benchmarks suggest this is at least kind of working (higher numbers are better):
Benchmark Mode Cnt Score Error Units
DecodingBenchmark.decodeFoosA thrpt 40 1219.959 ± 28.183 ops/s
DecodingBenchmark.decodeFoosC thrpt 40 1397.195 ± 15.402 ops/s
DecodingBenchmark.decodeIntsA thrpt 40 7538.238 ± 158.823 ops/s
DecodingBenchmark.decodeIntsC thrpt 40 7775.231 ± 121.456 ops/s
EncodingBenchmark.encodeFoosA thrpt 40 6186.539 ± 17.022 ops/s
EncodingBenchmark.encodeFoosC thrpt 40 6552.408 ± 127.446 ops/s
EncodingBenchmark.encodeIntsA thrpt 40 47262.384 ± 118.791 ops/s
EncodingBenchmark.encodeIntsC thrpt 40 92476.811 ± 367.129 ops/s
ParsingBenchmark.parseFoosA thrpt 40 2141.975 ± 13.181 ops/s
ParsingBenchmark.parseFoosC thrpt 40 3102.288 ± 19.494 ops/s
ParsingBenchmark.parseIntsA thrpt 40 10820.259 ± 38.952 ops/s
ParsingBenchmark.parseIntsC thrpt 40 33479.046 ± 94.416 ops/s
PrintingBenchmark.printFoosA thrpt 40 2873.823 ± 32.944 ops/s
PrintingBenchmark.printFoosC thrpt 40 3656.773 ± 9.829 ops/s
PrintingBenchmark.printIntsA thrpt 40 18641.201 ± 165.950 ops/s
PrintingBenchmark.printIntsC thrpt 40 22031.479 ± 80.269 ops/s
And allocation rates (lower is better):
Benchmark Cnt Score Error Units
DecodingBenchmark.decodeFoosA:gc.alloc.rate.norm 20 3129281.175 ± 2.273 B/op
DecodingBenchmark.decodeFoosC:gc.alloc.rate.norm 20 2745809.199 ± 57210.951 B/op
DecodingBenchmark.decodeIntsA:gc.alloc.rate.norm 20 599399.891 ± 21404.111 B/op
DecodingBenchmark.decodeIntsC:gc.alloc.rate.norm 20 528852.711 ± 5356.383 B/op
EncodingBenchmark.encodeFoosA:gc.alloc.rate.norm 20 526700.440 ± 1425.759 B/op
EncodingBenchmark.encodeFoosC:gc.alloc.rate.norm 20 429129.647 ± 1426.476 B/op
EncodingBenchmark.encodeIntsA:gc.alloc.rate.norm 20 88144.033 ± 7134.445 B/op
EncodingBenchmark.encodeIntsC:gc.alloc.rate.norm 20 48360.018 ± 0.035 B/op
ParsingBenchmark.parseFoosA:gc.alloc.rate.norm 20 1464616.780 ± 2138.191 B/op
ParsingBenchmark.parseFoosC:gc.alloc.rate.norm 20 735249.500 ± 2.852 B/op
ParsingBenchmark.parseIntsA:gc.alloc.rate.norm 20 326296.548 ± 1.540 B/op
ParsingBenchmark.parseIntsC:gc.alloc.rate.norm 20 105232.050 ± 0.096 B/op
PrintingBenchmark.printFoosA:gc.alloc.rate.norm 20 595216.561 ± 9978.237 B/op
PrintingBenchmark.printFoosC:gc.alloc.rate.norm 20 386712.468 ± 4190.873 B/op
PrintingBenchmark.printIntsA:gc.alloc.rate.norm 20 119592.091 ± 0.175 B/op
PrintingBenchmark.printIntsC:gc.alloc.rate.norm 20 95408.076 ± 0.147 B/op
The Foos
benchmarks work with a map containing case class values, and the Ints
ones are an array
of integers. C
suffixes indicate circe's throughput and A
is for Argonaut.
This section needs a lot of expanding.
circe uses Encoder
and Decoder
type classes for encoding and decoding. An Encoder[A]
instance
provides a function that will convert any A
to a JSON
, and a Decoder[A]
takes a Json
value
to either an exception or an A
. circe provides implicit instances of these type classes for many
types from the Scala standard library, including Int
, String
, and others. It also
provides instances for List[A]
, Option[A]
, and other generic types, but only if A
has an
Encoder
instance.
Suppose we have the following JSON document:
import io.circe._, io.circe.generic.auto._, io.circe.jawn._, io.circe.syntax._
import cats.data.Xor
val json: String = """
{
"id": "c730433b-082c-4984-9d66-855c243266f0",
"name": "Foo",
"counts": [1, 2, 3],
"values": {
"bar": true,
"baz": 100.001,
"qux": ["a", "b"]
}
}
"""
val doc: Json = parse(json).getOrElse(Json.empty)
In order to transform this document we need to create an HCursor
with the focus at the document's
root:
val cursor: HCursor = doc.hcursor
We can then use various operations to move the focus of the cursor around the document and to "modify" the current focus:
val reversedNameCursor: ACursor =
cursor.downField("name").withFocus(_.mapString(_.reverse))
We can then return to the root of the document and return its value with top
:
val reversedName: Option[Json] = reversedNameCursor.top
The result will contain the original document with the "name"
field reversed.
circe is a fork of Argonaut, and if you find it at all useful, you should thank Mark Hibberd, Tony Morris, Kenji Yoshida, and the rest of the Argonaut contributors.
circe is currently maintained by Travis Brown, Alexandre Archambault, and Vladimir Kostyukov. After the 0.3.0 release, all pull requests will require two sign-offs by a maintainer to be merged.
The circe project supports the Typelevel code of conduct and wants all of its channels (Gitter, GitHub, etc.) to be welcoming environments for everyone.
circe is licensed under the Apache License, Version 2.0 (the "License"); you may not use this software except in compliance with the License.
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.