Motivation

Considering e.g. a storage engine internal operation, its interaction with the SQL engine, or a database system as a whole we can see that row-like tuples are used in a variety of cases.

We already have IEP-54[1] that specifies a representation of table rows suitable for a key-value storage engine. It clearly separates key and value columns by keeping them in two distinct chunks. The separation is strictly defined by the table schema. The primary key columns go to the IEP-54[1] key chunk and other columns go to the value chunk.

However this approach does not cover all the cases when a database system deals with row-like data. For instance, execution of an SQL query produces a result set that may consist of rows containing columns from different tables, SQL-expressions, aggregate functions, etc. For such an arbitrary set of columns the distinction between key and value columns makes no sense at all.

When a storage engine deals with secondary indexes it picks some columns from a row and forms a tuple where the distinction between primary-key and non-primary-key columns does not play a role.

For a parametrized SQL query the query parameters also might need to be packed together into a single tuple to be fed to the database.

This proposal introduces a new binary format that might be suitable for all of these cases and more.

The design principles behind the format are as follows:

1. Keep it simple. Easy to build and parse.
2. Provide O(1) random access to every value.
3. Make it reasonably compact.

Description

We consider only the case when the tuple content is completely defined by a corresponding schema. A tuple schema may match a table schema (when tuples represent table rows) or may be very different (e.g. when tuples represent a result set of some complex SQL query).

Schema relieves tuples from bearing any metadata and thus provides space savings. The downside is that the tuple data can only be properly interpreted if the corresponding schema is available. For our use cases this is not a problem as the schema can always be determined from the context.

However, to provide fast and easy random access to individual tuple fields, the basic tuple layout depends only on the schema size. No other metadata is needed to get the required field. But one needs to learn the field data type from the schema when it is required to interpret the field as a value.

Tuple Layout

A binary tuple is kept as a sequence of fields contiguously going in memory one after another in the same order as they go in the tuple schema. The number of fields is defined by the schema too. Each field is a sequence of bytes that can be interpreted only after learning the field type in the schema.

The field location information is represented as a table of field offsets. This table allows to determine both the offset and the length of each field.

To be more precise the layout of a tuple looks like this:

1. Header;
2. Offset table;
3. Value area.

Header

It contains only one byte with flags.

Bits 0-1: Size class of the variable-length area.

Bit 2: Flag indicating that the size class is not optimal.

Size class encodes the number of bytes used for entries in the offset table. It is encoded like this:

0b00 - 1 byte

0b01 - 2 bytes

0b10 - 4 bytes

0b11 - 8 bytes

So the entry size can be calculated as follows:

Size = 1 << (flags & 0b11)

NOTE: By the way, the last option works fine for C++ and possibly other languages but for Java it will be hard to support, so it might be excluded.

Offset Table

The number of elements in the table is equal to the number of columns in the schema.

The size of table elements might be 1, 2, 4, or 8 bytes depending on the header flags.

Each element in the table specifies where the corresponding field ends. In other words, for each given field the table stores the offset where the next field starts. The last element in the table specifies the end of the last field and at the same time the end of the whole value area and consequently the end of the entire tuple.

The offset is measured relative to the start of the value area.

The start of the first element is implied and is always equal to zero.

Value area

A value area contains data for each tuple field one after another as described above. Potentially a field can contain any kind of data.  However there are rules for representation of common data types there. These rules are specified in the following section. Data types for which there are no specified rules still may be put into a tuple as a Binary field, but the interpretation in such cases is fully up to a particular application.

Field Encoding

A tuple field is a sequence of bytes. This sequence is interpreted according to the associated data type provided by the tuple schema. There are different encoding rules for different values. For some data types the rules include a very simple compression mechanism. So even if the original type has fixed size it may take different number of bytes when encoded as a tuple field.

If a value is equal to NULL then it is absent in the value area. This means that in the offset table the corresponding entry is equal to the previous entry.

For some variable-length types we can encounter a value with zero length. To distinguish a zero-length variable-length value from a null value we use a special magic byte that denotes an empty value. That is an empty value is encoded as single-byte sequence – 0x80. In turn if the original variable-length value starts with the magic byte 0x80 byte then this byte is doubled.

The Number and Decimal types are never empty, at least one significant byte is always present. The variable-length types that use the magic byte for encoding are as follows:

• String;
• Binary;
• Bitmask.

The list of supported data types is as follows:

Integer Representation

All integer values are stored in the little-endian byte order.

For integer values it is possible to keep only their significant bytes, omitting high insignificant bytes. That is even if the original data type of a field is INTEGER and should occupy 4 bytes but the actual value is below 128  and above -129  then we can store it just as a single byte and reflect this in the offset table.

For signed integers the most significant bit is the sign bit. Thus a compressed signed integer must be sign-extended on decompression.

In SQL standard there are no unsigned integers. So their support is omitted for now. But should it become needed for some reason then a compressed unsigned integer must be zero-extended on decompression.

UUID Representation

A UUID value occupies 16 bytes. It is encoded with a slight twist. It is split into two 8-byte integers – the most significant and least significant ones. The most significant goes first, the least significant goes after it. But each of these integers is stored in little-endian byte order. So a UUID 00112233-4455-6677-8899-aabbccddeeff will be encoded as byte sequence 77 66 55 44 33 22 11 00 ff ee dd cc bb aa 99 88.

Date Representation

A date field occupies 3 bytes.

Time Representation

A time field may occupy 4, 5, or 6 bytes depending on the factional part precision that may  be up to a millisecond, microsecond,  and nanosecond respectively.

DateTime Representation

A DateTime field is represented as a Date followed by a Time.  So it may occupy from 7 to 9 bytes depending on the factional part of the Time value.

Timestamp Representation

A timestamp field may occupy either 8 or 12 bytes. It consists of a 64-bit seconds part optionally followed by a 32-bit nanoseconds part.  The seconds part is a signed number representing the time since so called Epoch time – Jan 1, 1970 00:00:00. The nanoseconds part varies between 0 and 999,999,999. It can be omitted meaning that it is equal to zero.

Bitmask Representation

A bitmask field is represented as a byte sequence in little-endian order.

Number and Decimal representation

Number and Decimal have identical representation. The Decimal type has a scale parameter but it is stored in the schema rather than in each tuple. Fields of these types are represented as a byte sequence containing the two's-complement binary value. Unlike all the other types here the bytes go in big-endian order. That is the most significant byte is at the very start of the sequence and the least significant byte is at its end.

Duration Representation

A duration field may occupy either 8 or 12 bytes. It consists of a 64-bit signed seconds part optionally followed by a 32-bit nanoseconds part in range from 0 to 999,999,999.

Period Representation

A period field may occupy 3, 6, or 12 bytes. It consists of 3 signed varlen int parts: years, months, days. When every part fits into 8 bits, we write 3 signed 8-bit integers. When every part fits into 16 bits, we write 3 signed 16-bit integers. Otherwise, every part is written as a signed 32-bit integer.

Unlike Date, every part is independent from others and can be in range from Integer.MIN_VALUE to Integer.MAX_VALUE.

Boolean Representation

A single byte containing 1 for the value of true and 0 for the value of false.

Corollary

If the number of fields is N and t is an array that stores a binary tuple we can find the answers for the following:

How many bytes are occupied by one offset table entry?

offsetEntryBytes = 1 << (t[0] & 0b11);

How many bytes are occupied by the offset table?

offsetTableBytes = N * offsetEntryBytes;

What is offset of the value area?

valueBaseOffset = 1 + offsetTableBytes;

What is the whole tuple size?

tupleSize = valueBaseOffset + valueAreaSize;   // valueAreaSize goes from the last offset table entry

Corollary #2

In order to build a tuple using minimum possible space it is required to learn what is the total length of all non-null values. After that we can figure out the minimum possible size of the offset table entries.

Thus, generally speaking, building a binary tuple is a two-pass procedure. Sometimes it might be possible to turn this into a single pass (almost) by over-provisioning the allocated storage for the worst case and then fixing it up at the end.

Reference Links

1. IEP-54: Schema-first Approach

Tickets