Eminim Save

JSON serialization framework for Nim, works from a Stream directly to any type and back. Depends only on stdlib.

Project README

Eminim — JSON serialization framework for Nim

About

This package provides the jsonTo, loadJson procs and jsonItems iterator which deserializes the specified type from a Stream. The storeJson procs are used to write the JSON representation of a location into a Stream. Low level initFromJson and storeJson procs can be overloaded, in order to support arbitary container types, i.e. jsmartptrs.nim.

Usage

import std/streams, eminim

type
  Foo = ref object
    value: int
    next: Foo

let d = Foo(value: 1, next: Foo(value: 2, next: nil))
var s = newStringStream()
# Make a roundtrip
s.storeJson(d) # writes JSON from a location
# Stream's data now contain:
# {"value":1,"next":{"value":2,"next":null}}
s.setPosition(0)
let a = s.jsonTo(Foo) # reads JSON and transform to a type
# Alternatively load directly into a location
s = newStringStream()
s.storeJson(d)
s.setPosition(0)
var b: Foo
s.loadJson(b)

Features

  • Serializing and deserializing directly into Stream. For common usage it is done automatically. Generally speaking intervation is needed when working with ptr types.
  • Supports options, sets and tables by default.
  • Uses nim identifier equality algorithm to compare JSON fields. Which means fields written in camelCase or snake_case are equal.
  • Overloading serialization procs. See Examples
  • Strict field checking can be disabled at compile-time with -d:emiLenient. Meaning you can parse complex JSON structures like the World Bank dataset and retrieve only the fields you're interested.

How it works

It generates code, in compile time, to use directly the JsonParser, without creating an intermediate JsonNode tree.

For example:

type
  Bar = object
    case kind: Fruit
    of Banana:
      bad: float
      banana: int
    of Apple: apple: string

let s = newStringStream("""{"kind":"Apple","apple":"world"}""")
let a = s.jsonTo(Bar)

Produces this code:

proc initFromJson(dst: var Bar, p: var JsonParser) =
  eat(p, tkCurlyLe)
  while p.tok != tkCurlyRi:
    if p.tok != tkString:
      raiseParseErr(p, "string literal as key")
    case p.a
    of "kind":
      discard getTok(p)
      eat(p, tkColon)
      var kindTmp`gensym0: Fruit
      initFromJson(kindTmp`gensym0, p)
      dst = (typeof dst)(kind: kindTmp`gensym0)
      if p.tok != tkComma:
        break
      discard getTok(p)
      while p.tok != tkCurlyRi:
        if p.tok != tkString:
          raiseParseErr(p, "string literal as key")
        case dst.kind
        of Banana:
          case p.a
          of "bad":
            discard getTok(p)
            eat(p, tkColon)
            initFromJson(dst.bad, p)
          of "banana":
            discard getTok(p)
            eat(p, tkColon)
            initFromJson(dst.banana, p)
          else:
            raiseParseErr(p, "valid object field")
        of Apple:
          case p.a
          of "apple":
            discard getTok(p)
            eat(p, tkColon)
            initFromJson(dst.apple, p)
          else:
            raiseParseErr(p, "valid object field")
        if p.tok != tkComma:
          break
        discard getTok(p)
    else:
      raiseParseErr(p, "valid object field")
    if p.tok != tkComma:
      break
    discard getTok(p)
  eat(p, tkCurlyRi)

proc jsonTo(s: Stream, t: typedesc[Bar]): Bar =
  var p: JsonParser
  open(p, s, "unknown file")
  try:
    discard getTok(p)
    initFromJson(result, p)
    eat(p, tkEof)
  finally:
    close(p)

The jsonItems iterator

Inspired by the Machine Learning over JSON article, I originally designed eminim as a way to parse JSON datasets directly into Tensors. It's still very capable in this application.

type
  IrisPlant = object
    sepalLength: float32
    sepalWidth: float32
    petalLength: float32
    petalWidth: float32
    species: string

let fs = newFileStream("iris.json")
for x in jsonItems(fs, IrisPlant):
  # you have read an item from the iris dataset,
  # use it to create a Tensor

Limitations

  • Limited support of object variants. The discriminant field is expected first. Also there can be no fields before and after the case section. In all other cases it fails with a JsonParserError. This limitation is hard to improve. The current state might fit some use-cases and it's better than nothing.

  • Borrowing proc initFromJson[T](dst: var T; p: var JsonParser) for distinct types isn't currently working. Blocked by a Nim bug. Use overloads for now. Or you can easily override this behaviour by copying these lines in your project:

    from typetraits import distinctBase
    proc storeJson*[T: distinct](s: Stream; x: T) = storeJson(s, x.distinctBase)
    proc initFromJson*[T: distinct](dst: var T; p: var JsonParser) = initFromJson(dst.distinctBase, p)
    
  • Custom pragmas are not supported. Unless hasCustomPragma improves, this feature won't be added. You can currently substitute skipped fields by creating empty overloads.

Binary serialization

If you need binary serialization instead, check planetis-m/bingo.

Acknowledgements

  • Thanks to @krux02 for his review and valuable feedback. This rewrite wouldn't be possible without his work on json.to.
  • treeform, for his "not-strict" idea, also check out his package treeform/jsony.
Open Source Agenda is not affiliated with "Eminim" Project. README Source: planetis-m/eminim