*** DRAFT ***
This document describes SQLite's JSONB binary encoding of JSON.
Beginning with version 3.45.0 (pending), SQLite supports an alternative binary encoding of JSON which we call "JSONB". JSONB is a binary format that stored as a BLOB.
The advantage of JSONB over ordinary text RFC 8259 JSON is that JSONB is both slightly smaller (by between 5% and 10% in most cases) and can be processed in less than half the number of CPU cycles. The built-in JSON SQL functions of SQLite can accept either ordinary text JSON or the binary JSONB encoding for any of their JSON inputs.
The "JSONB" name is inspired by PostgreSQL, but the on-disk format for SQLite's JSONB is not the same as PostgreSQL's. The two formats have the same name, but they have wildly different internal representations and are not in any way binary compatible.
The central idea behind this JSONB specification is that each element begins with a header that includes the size and type of that element. The header takes the place of punctuation such as double-quotes, curly-brackes, square-brackets, commas, and colons. Since the size and type of each element is contained in its header, the element can be read faster since it is no longer necessary to carefully scan forward looking for the closing delimiter. The payload of JSONB is the same as for corresponding text JSON. The same payload bytes occur in the same order. The only real difference between JSONB and ordinary text JSON is that JSONB includes a binary header on each element and omits delimiter and separator punctuation.
JSONB is not intended as an external format to be used by applications. JSONB is designed for internal use by SQLite only. Programmers do not need to understand the JSONB format in order to use it effectively. Applications should access JSONB only through the JSON SQL functions, not by looking at individual bytes of the BLOB.
However, JSONB is intended to be portable and backwards compatible for all future versions of SQLite. In other words, you should not have to export and reimport your SQLite database files when you upgrade to a newer SQLite version. For that reason, the JSONB format needs to be well-defined.
This document is therefore similar in purpose to the SQLite database file format document that describes the on-disk format of an SQLite database file. Applications are not expected to directly read and write the bits and bytes of SQLite database files. The SQLite database file format is carefully documented so that it can be stable and enduring. In the same way, the JSONB representation of JSON is documented here so that it too can be stable and enduring, not so that applications can read or writes individual bytes.
JSONB is a direct translation of the underlying text JSON. The difference is that JSONB uses a binary encoding that is faster to parse compared to the detailed syntax of text JSON.
Each JSON element is encoded as a header and a payload. The header determines type of element (string, numeric, boolean, null, object, or array) and the size of the payload. The header can be between 1 and 9 bytes in size. The payload can be any size from zero bytes up to the maximum allowed BLOB size.
The upper four bits of the first byte of the header determine size of the header and possibly also the size of the payload. If the upper four bits have a value between 0 and 11, then the header is exactly one byte in size and the payload size is determined by those upper four bits. If the upper four bits have a value between 12 and 15, that means that the total header size is 2, 3, 5, or 9 bytes and the payload size is unsigned big-endian integer that is contained in the subsequent bytes. The size integer is the one byte that following the initial header byte if the upper four bits are 12, two bytes if the upper bits are 13, four bytes if the upper bits are 14, and eight bytes if the upper bits are 15. The current design of SQLite does not support BLOB values larger than 2GiB, so the eight-byte variant of the payload size integer will never be used by the current code. The eight-byte payload size integer is included in the specification to allow for future expansion.
The header for an element does not need to be in its simplest form. For example, consider the JSON numeric value "1". That element can be encode in five different ways:
The shortest encoding is preferred, of course, and usually happens with primitive elements such as numbers. However the total size of an array or object might not be known exactly when the header of the element is first generated. It is convenient to reserve space for the largest possible header and then go back and fill in the correct payload size at the end. This technique can result in array or object headers that are larger than absolutely necessary.
The least-significant four bits of the first byte of the header (the first byte masked against 0x0f) determine element type. The following codes are used:
NULL → The element is a JSON "null". The payload size for a true JSON NULL must must be zero. Future versions of SQLite might extend the JSONB format with elements that have a zero element type but a non-zero size. In that way, legacy versions of SQLite will interpret the element as a NULL for backwards compatibility while newer versions will interpret the element in some other way.
TRUE → The element is a JSON "true". The payload size must be zero for a actual "true" value. Elements with type 1 and a non-zero payload size are reserved for future expansion. Legacy implementations that see an element type of 1 with a non-zero payload size should continue to interpret that element as "true" for compatibility.
FALSE → The element is a JSON "false". The payload size must be zero for a actual "false" value. Elements with type 2 and a non-zero payload size are reserved for future expansion. Legacy implementations that see an element type of 2 with a non-zero payload size should continue to interpret that element as "false" for compatibility.
INT → The element is a JSON integer value in the canonical RFC 8259 format, without extensions. The payload is the ASCII text representation of that numeric value.
INT5 → The element is a JSON integer value that is not in the canonical format. The payload is the ASCII text representation of that numeric value. Because the payload is in a non-standard format, it will need to be translated when the JSONB is converted into RFC 8259 text JSON.
FLOAT → The element is a JSON floating-point value in the canonical RFC 8259 format, without extensions. The payload is the ASCII text representation of that numeric value.
FLOAT5 → The element is a JSON floating-point value that is not in the canonical format. The payload is the ASCII text representation of that numeric value. Because the payload is in a non-standard format, it will need to be translated when the JSONB is converted into RFC 8259 text JSON.
TEXT → The element is a JSON string value that does not contain any escapes nor any characters that need to be escaped for either SQL or JSON. The payload is the UTF8 text representation of the string value. The payload does not include string delimiters.
TEXTJ → The element is a JSON string value that contains RFC 8259 character escapes (such as "\n" or "\u0020"). Those escapes will need to be translated into actual UTF8 if this element is extracted into SQL. The payload is the UTF8 text representation of the escaped string value. The payload does not include string delimiters.
TEXT5 → The element is a JSON string value that contains character escapes, including some character escapes that part of JSON5 and which are not found in the canonical RFC 8259 spec. Those escapes will need to be translated into standard JSON prior to rendering the JSON as text, or into their actual UTF8 characters if this element is extracted into SQL. The payload is the UTF8 text representation of the escaped string value. The payload does not include string delimiters.
TEXTRAW → The element is a JSON string value that contains UTF8 characters that need to be escaped if this string is rendered into standard JSON text. The payload does not include string delimiters.
ARRAY → The element is a JSON array. The payload contains JSONB elements that comprise values contained within the array.
OBJECT → The element is a JSON object. The payload contains pairs of JSONB elements that comprise entries for the JSON object. The first element in each pair must be a string (types 7 through 10). The second element of each pair may be any types, including nested arrays or objects.
RESERVED-13 → Reserved for future expansion. Legacy implements that encounter this element type should raise an error.
RESERVED-14 → Reserved for future expansion. Legacy implements that encounter this element type should raise an error.
RESERVED-15 → Reserved for future expansion. Legacy implements that encounter this element type should raise an error.
Element types outside the range of 0 to 12 are reserved for future expansion. The current implement raises an error if see an element type other than those listed above. However, future versions of SQLite might use of the three remaining element types to implement indexing or similar optimizations, to speed up lookup against large JSON arrays and/or objects.
A key goal of JSONB is that it should be quick to translate to and from text JSON and/or be constructed from SQL values. When converting from text into JSONB, we do not want the converter subroutine to burn CPU cycles converting elements values into some standard format which might never be used. Format conversion is "lazy" - it is deferred until actually needed. This has implications for the JSONB format design:
Numeric values are stored as text, not a numbers. The values are a direct copy of the text JSON values from which they are derived.
There are multiple element types depending on the details of value formats. For example, INT is used for pure RFC-8259 integer literals and INT5 exists for JSON5 extensions such as hexadecimal notation. FLOAT is used for pure RFC-8259 floating point literals and FLOAT5 is used for JSON5 extensions. There are four different representations of strings, depending on where the string came from and how special characters within the string are escaped.
A second goal of JSONB is that it should be capable of serving as the "parse tree" for JSON when a JSON value is being processed by the various JSON SQL functions built into SQLite. Before JSONB was developed, operations such json_replace() and json_patch() and similar worked in three stages:
JSONB seeks to serve as the internal format directly - bypassing the first and third stages of that process. Since most of the CPU cycles are spent on the first and third stages, that suggests that JSONB processing will be much faster than text JSON processing.
(Aside: As of this writing, the legacy internal binary format is still used for many operations. Only json_extract() has, thus far, been converted to use JSONB directly. The plan is to convert all JSON processing to use JSONB internally. Hopefully this will be accomplished prior to the first official release.)
So when processing JSONB, only the second stage of the three-stage process is required. But when processing text JSON, it is still necessary to do stages one and three. If JSONB is to be used as the internal binary representation, this is yet another reason to store numeric values as text. Storing numbers as text minimizes the amount of conversion work needed for stages one and three. This is also why there are four different representations of text in JSONB. Different text representations are used for text coming from different sources (RFC-8259 JSON, JSON5, or SQL string values) and conversions only happen if and when they are actually needed.
A valid JSONB BLOB consists of a single JSON element. The element must exactly fill the BLOB. This one element is often a JSON object or array and those usually contain additional elements as its payload, but the element can be a primite value such a string, number, boolean, or null.
When the built-in JSON functions are attempting to determine if a BLOB argument is a JSONB or just a random BLOB, they look at the header of the outer element to see that it is well-formed and that the element completely fills the BLOB. If these conditions are met, then the BLOB is accepted as a JSONB value.
This page last modified on 2023-12-05 19:57:45 UTC
*** DRAFT ***