Find file
e78015a Jan 9, 2017
143 lines (112 sloc) 5.61 KB

Specification of the Compact Message Format (CMF).

CMF is a binary format that combines the speed and compact size of binary formats and combines it with provably correct markup which became well known with formats like XML. CMF also introduces type-safety to avoid a number being interpreted as a string, or vice-versa.

In CMF we prioritize compact messages over speed and this means that all numbers are stored as variable-width numbers. Parsing of this is still magnitudes faster than parsing a string representation, so this should not be understood to mean that CMF is slow.

CMF is in essence a flat list of tokens. Each token is comprised of up to 3 elements. The name, the format and the value. When a format is "BoolTrue" then the value can be omitted.

Each CMF-token is completely self-contained and even a reader that doesn't know the schema of the message type it is parsing can still isolate and iterate individual tokens and extract their values. This makes the format exceptionally useful for extensibility because a reader or an old schema can just skip over unknown tokens.

At this point it may be useful to compare CMF to XML. XML is rather more complex and rich, but if we focus only on the elements the comparison can be meaningful to understand the ideas behind CMF.

In XML you'd be able to write something like this;


In XML everything is text (utf8) whereas in CMF we use a binary format for everything. One implication of this is that we need a translation table to map the element names (like 'population') to CMF names which are essentially just numbers and coded in an enumeration. Much like XML the names of a token can be reused in a different context so you won't run out of them.

Second difference is that CMF is a flat list, which means that you can use closing-tags, but there really is no need in most cases. Exception may be when the section between the open and close tags is meant to change namespace.

Consider a simple mapping table for our XML example.

  1. city
  2. name
  3. name:en
  4. founded
  5. population

The result is a 24 byte message comprising of a list of tokens (one per row) like this:

Name Format Value Bytes
city bool true 0C
name string Köln 12 05 4B C3 B6 6C 6E
name:en string Cologne 13 07 43 6F 6C 6F 67 6E 65
founded NegativeNumber -38 21 26
population PositiveNumber 1060584 05 BF DC 68


In the CMF we identified individual tokens. Each token has up to 3 items; a name, a format and a value.

The name and format are encoded in a single step and the value is encoded based on which format is chosen.

The name and format are merged into one byte, where possible. The division is simple where the format takes the 3 least significant bits and the name takes the rest. In case the name is an enum that has a value of 31 or higher we put the magic value of 31 (highest value possible in a 5 bit number) in this byte and then follow up with the name in subsequent bytes. The designer of the schema is encouraged to use a different scope for each type of message in order to maximize reuse of the name values and avoid using more than one byte. Long experience shows that with well chosen context-divisions the 30 names is more than enough.

The 3 bits format expands to a list of 7 different formats, listed in the table below. The format specified for a token decides how the value is encoded in the stream.

Format enum-value Value-encoding
PositiveNumber 0 Followed by a var-int encoded integer
NegativeNumber 1 The value is multiplied with -1 and then serialized exactly like PositiveNumber
String 2 var-int Length (in bytes) first, then UTF8 encoded string
ByteArray 3 var-int Length first, then the actual bytes
BoolTrue 4
BoolFalse 5
Double 6 The little-endian format of a native double in exactly 8 bytes

Variable-width-integer encoding (var-int)

The Var-Int format is a simple variable-byte encoding which Bitcoin has used for years. Because there is no stand-alone specification on this format itself I'll include a terse one in here. The simplest explanation is that it splits the entire number in blocks of 7 bits and serializes it lowest-byte-first into the stream reserving the highest (8th) bit as a flag to indicate if more bytes will come. Additionally, when a bit is used to indicate another byte this will decrease the value encoded by one.


Input Out bytes
0X7F 0X7F
0X80 0X80 0X00
0XFF 0X80 0X7F
0X407F 0XFF 0X7F
0X4080 0X80 0X80 0X00

Details for the Example

In the introduction there is an example message about a city Cologne that is first described in XML and then encoded into CMF with byte-values listed in the table. To help readers understand the format better, here is the step by step logic towards the actual byte values that make up our example message.

City is an enum of 1. The 'bool-true' is a value of 4. Which makes binary 00001 100 which is 0x0C.

The second token 'Name' is an enum of 2. The 'string' is a value of 2. Which makes binary 00010 010 which is 0x12 The string will be encoded in UTF-8. Which serializes the 4 character German name "Köln" to 5 bytes. This is the next byte; 5. Then the actual string is added.

Founded is en enum of 4. With the format of Negative number this is binary 000100 001 which is 0x21. The Negative Number is essentially a var-int like PositiveNumber. Just without the minus. So the next value is 0x26.

Population is an enum of 5. The PositiveNumber is zero. This is 00101 000 which is 0x05.