A fast and minimal JSON parser and transformer that works on unstructured JSON. It works by parsing input and calling the given callback function when encountering each item.

Read more on the introduction article.

Sometimes we just want to make some minimal changes to a JSON document, or do some generic transformations without fully unmarshalling it. For example, removing blacklist fields from response JSON. Why spend all the cost on unmarshalling into a map[string]interface{} just to immediate marshal it again.

The following code is taken from StackOverflow:

{
  "responseHeader": {
    "status": 0,
    "QTime": 0,
    "params": {
      "q": "solo",
      "wt": "json"
    }
  },
  "response": {
    "numFound": 2,
    "start": 0,
    "docs": [
      { "name": "foo" },
      { "name": "bar" }
    ]
  }
}

With µjson, we can quickly write a simple transformation to remove "responseHeader" completely from all responses, once and forever:

func removeBlacklistFields(input []byte) []byte {
    blacklistFields := [][]byte{
		[]byte(`"responseHeader"`), // note the quotes
	}
	b := make([]byte, 0, 1024)
	err := ujson.Walk(input, func(_ int, key, value []byte) bool {
		if len(key) != 0 {
			for _, blacklist := range blacklistFields {
				if bytes.Equal(key, blacklist) {
					// remove the key and value from the output
					return false
				}
			}
		}
		// write to output
		if len(b) != 0 && ujson.ShouldAddComma(value, b[len(b)-1]) {
			b = append(b, ',')
		}
		if len(key) > 0 {
			b = append(b, key...)
			b = append(b, ':')
		}
		b = append(b, value...)
		return true
	})
	if err != nil {
		panic(err)
	}
	return b
 }

The original scenario that leads me to write the package is because of int64. When working in Go and PostgreSQL, I use int64 (instead of string) for ids because it’s more effective and has enormous space for randomly generated ids. It’s not as big as UUID, 128 bits, but still big enough for production use. In PostgreSQL, those ids can be stored as bigint and being effectively indexed. But for JavaScript, it can only process integer up to 53 bits (JavaScript has BigInt but that’s a different story, and using it will make things even more complicated).

So we need to wrap those int64s into strings before sending them to JavaScript. In Go and PostgreSQL, the JSON is {"order_id": 12345678} but JavaScript will see it as {"order_id": "12345678"} (note that the value is quoted). In Go, we can define a custom type and implement the json.Marshaler interface. But in PostgreSQL, that’s just not possible or too complicated. I wrote a service that receives JSON from PostgreSQL and converts it to be consumable by JavaScript. The service also removes some blacklisted keys or does some other transformations (for example, change orderId to order_id).

1. Walk through unstructured JSON:

   • Print all keys and values
   • Extract some values

2. Transform unstructured JSON:

   • Remove all spaces
   • Reformat
   • Remove blacklist fields
   • Wrap int64 in string for processing by JavaScript

without fully unmarshalling it into a map[string]interface{}.

See usage and examples on pkg.go.dev.

Important: Behaviour is undefined on invalid JSON. Use on trusted input only. For untrusted input, you may want to run it through json.Valid() first.

The single most important function is Walk(input, callback), which parses the input JSON and call callback function for each key/value pair processed.

The callback function is called when an object key/value or an array key is encountered. It receives 3 params in order: level, key and value.

• level is the indentation level of the JSON, if you format it properly. It starts from 0. It increases after entering an object or array and decreases after leaving.

• key is the raw key of the current object or empty otherwise. It can be a double-quoted string or empty.

• value is the raw value of the current item or a bracket. It can be a string, number, boolean, null, or one of the following brackets: { } [ ].

It’s important to note that key and value are provided as raw. Strings are always double-quoted. It’s there for keeping the library fast and ignoring unnecessary operations. For example, when you only want to reformat the output JSON properly; you don’t want to unquote those strings and then immediately quote them again; you just need to output them unmodified. And there are ujson.Unquote() and ujson.AppendQuote() when you need to get the original strings.

For valid JSON, values will never be empty. We can test the first byte of value (value[0]) to get its type:

• n: Null (null)
• f, t: Boolean (false, true)
• 0-9, -: Number
• ": String, see Unquote()
• [, ]: Array
• {, }: Object

When processing arrays and objects, first the open bracket ([, {) will be provided as value, followed by its children, and finally the close bracket (], }). When encountering open brackets, You can make the callback function return false to skip the array/object entirely.

This example gives a quick idea about how µjson works.

 input := []byte(`{  
     "id": 12345,    
     "name": "foo",  
     "numbers": ["one", "two"],  
     "tags": {"color": "red", "priority": "high"},   
     "active": true  
 }`) 
 ujson.Walk(input, func(level int, key, value []byte) bool { 
     fmt.Printf("%2v% 12s : %s\n", level, key, value)    
     return true 
 })

{
   "id": 12345,
   "name": "foo",
   "numbers": ["one", "two"],
   "tags": {"color": "red", "priority": "high"},
   "active": true
}

Calling Walk() with the above input will produce:

To easily get an idea on level, key and value, here are the simplest examples:

 input0 := []byte(`true`)
 ujson.Walk(input0, func(level int, key, value []byte) bool {
     fmt.Printf("level=%v key=%s value=%s\n", level, key, value)
     return true
 })
 // output:
 //   level=0 key= value=true

 input1 := []byte(`{ "key": 42 }`)
 ujson.Walk(input1, func(level int, key, value []byte) bool {
     fmt.Printf("level=%v key=%s value=%s\n", level, key, value)
     return true
 })
 // output:
 //   level=0 key= value={
 //   level=1 key="key" value=42
 //   level=0 key= value=}

 input2 := []byte(`[ true ]`)
 ujson.Walk(input2, func(level int, key, value []byte) bool {
     fmt.Printf("level=%v key=%s value=%s\n", level, key, value)
     return true
 })
 // output:
 //   level=0 key= value=[
 //   level=1 key= value=true
 //   level=0 key= value=]

In the first example, there is only a single boolean value. The callback function is called once with level=0, key is empty and value=true.

In the second example, the callback function is called 3 times. Two times for open and close brackets with level=0, key is empty and value is the bracketed character. The other time for the only key with level=1, key is "key" and value=42. Note that the key is quoted and you need to call ujson.Unquote() to retrieve the unquoted string.

The last example is like the second, but with an array instead. Keys are always empty inside arrays.

In this example, the input JSON is formatted with correct indentation. As processing the input key by key, the callback function reconstructs the JSON. It outputs each key/value pair in its own line, prefixed with spaces equal to the param level. There is a catch, though. Valid JSON requires commas between values in objects and arrays. So there is ujson.ShouldAddComma() for checking whether a comma should be inserted.

 input := []byte(`{"id":12345,"name":"foo","numbers":["one","two"],"tags":{"color":"red","priority":"high"},"active":true}`)

 b := make([]byte, 0, 1024)
 err := ujson.Walk(input, func(level int, key, value []byte) bool {
     if len(b) != 0 && ujson.ShouldAddComma(value, b[len(b)-1]) {
         b = append(b, ',')
     }
     b = append(b, '\n')
     for i := 0; i < level; i++ {
         b = append(b, '\t')
     }
     if len(key) > 0 {
         b = append(b, key...)
         b = append(b, `: `...)
     }
     b = append(b, value...)
     return true
 })
 if err != nil {
     panic(err)
 }
 fmt.Printf("%s\n", b)

{
	"id": 12345,
	"name": "foo",
	"numbers": [
		"one",
		"two"
	],
	"tags": {
		"color": "red",
		"priority": "high"
	},
	"active": true
}

There is a built-in method ujson.Reconstruct() when you want to remove all the whitespaces.

This example demonstrates removing some keys from the input JSON. The key param is compared with a pre-defined list. If there is a match, the blacklisted key and its value are dropped. The callback function returns false for skipping the entire value (which may be an object or array). Note that the list is quoted, i.e. "numbers" and "active" instead of number and active. For more advanced checking, you may want to run ujson.Unquote() on the key.

 input := []byte(`{
     "id": 12345,
     "name": "foo",
     "numbers": ["one", "two"],
     "tags": {"color": "red", "priority": "high"},
     "active": true
 }`)

 blacklistFields := [][]byte{
     []byte(`"numbers"`), // note the quotes
     []byte(`"active"`),
 }
 b := make([]byte, 0, 1024)
 err := ujson.Walk(input, func(_ int, key, value []byte) bool {
     for _, blacklist := range blacklistFields {
         if bytes.Equal(key, blacklist) {
             // remove the key and value from the output
             return false
         }
     }

     // write to output
     if len(b) != 0 && ujson.ShouldAddComma(value, b[len(b)-1]) {
         b = append(b, ',')
     }
     if len(key) > 0 {
         b = append(b, key...)
         b = append(b, ':')
     }
     b = append(b, value...)
     return true
 })
 if err != nil {
     panic(err)
 }
 fmt.Printf("%s\n", b)

{"id":12345,"name":"foo","tags":{"color":"red","priority":"high"}}

This is the original motivation behind µjson. The following example finds keys ending with _id" ("order_id", "item_id", etc.) and converts their values from numbers to strings, by simply wrapping them in double-quotes.

For valid JSON, values will never be empty. We can test the first byte of value (value[0]) to get its type:

• n: Null (null)
• f, t: Boolean (false, true)
• 0-9, -: Number
• ": String, see Unquote()
• [, ]: Array
• {, }: Object

In this case, we check value[0] within 0…9 to see whether it’s a number, then insert double-quotes.

 input := []byte(`{"order_id": 12345678901234, "number": 12, "item_id": 12345678905678, "counting": [1,"2",3]}`)

 suffix := []byte(`_id"`) // note the ending quote "
 b := make([]byte, 0, 256)
 err := ujson.Walk(input, func(_ int, key, value []byte) bool {
     // Test for keys with suffix _id" and value is an int64 number. For valid json,
     // values will never be empty, so we can safely test only the first byte.
     shouldWrap := bytes.HasSuffix(key, suffix) && value[0] > '0' && value[0] <= '9'

     // transform the input, wrap values in double quotes
     if len(b) != 0 && ujson.ShouldAddComma(value, b[len(b)-1]) {
         b = append(b, ',')
     }
     if len(key) > 0 {
         b = append(b, key...)
         b = append(b, ':')
     }
     if shouldWrap {
         b = append(b, '"')
     }
     b = append(b, value...)
     if shouldWrap {
         b = append(b, '"')
     }
     return true
 })
 if err != nil {
     panic(err)
 }
 fmt.Printf("%s\n", b)

{"order_id":"12345678901234","number":12,"item_id":"12345678905678","counting":[1,"2",3]}

After processing, the numbers in "order_id" and "item_id" are quoted as strings. And JavaScript should be happy now! 🎉 🎉

MIT