Array Object

Self-describing binary format for arrays of integers, real numbers, complex numbers and strings, designed for object storage, database and single file.

ArrayObject is a part of dbgbb project.

Highlights

• The data is self-describing and can inflate itself into typed variables.
• No nested structures, no tuple, no dataset name, always a simple array of uniform data.
• Generic integer and float types absorb the difference of type sizes.
• Automatic compression using variable length integer/float and dictionary-coder for string.
• The data is stored in the minimal data size.
• Conversions from/into ndarray and nalgebra are supported.

Examples

Encoding and decording:

use array_object::*;

fn main() {
    // Convert data into binary
    let original = vec![1u32, 2, 3, 4];
    let obj: ArrayObject = original.clone().try_into().unwrap();
    let packed = obj.pack(); // This converts the data into Vec<u8>.

    // Restore data
    let unpacked = ArrayObject::unpack(packed).unwrap();
    let inflated: Vec<u32> = unpacked.try_into().unwrap();
    assert_eq!(original, inflated);
}

One can also use the macros to write and read a file:

use array_object::*;

fn main() {
    // Save into a file
    let original = vec![1f64, 2.2, -1.1, 5.6];
    export_obj!("testdata.bin", original.clone()); // The type has to be known at this point.

    // Load from a file
    let restored: Vec<f64> = import_obj!("testdata.bin"); // The type annotation is required.
    assert_eq!(original, restored);
}

Crate Features

Format

The data format is automatically selected to minimize the datasize.

Integer

Integer is either unsigned or signed, which is determined when [ArrayObject] is constructed. The zigzag encoding is used for signed integers. When restored to a variable, data is automatically converted into the desired integer type if the ranges overlap.

Scalar

• Short Integer (5bit)
  The data is stored in the same byte as the footer. Thus the total data size is only one byte.

• Variable Length (8 x n bit)
  The integer is shortened to 8bit x (smallest number).

Array

• Fixed Length (8bit, 16bit, 32bit, 64bit, 128bit)
  Use the smallest possible size. All the elements have the same size.
• Variable Length (8bit, 16bit, 32bit, 63bit, 64-128bit variable)
  The integer is shortened to the smallest possible size. Each four integers, one byte is added to indicate the size of each integer type. If the integer is longer than 63 bit, one byte is added to indicate how many bytes should be read additionally.

Float (Real, Complex)

Currently 32bit and 64bit floating numbers are supported.

Scalar

• Fixed Length (16bit, 32bit, 64bit, 128bit)
  Use the smallest possible size without loss of precision.

Array

• Fixed Length (16bit, 32bit, 64bit, 128bit)
  Use the smallest possible size without loss of precision. All the numbers have the same size.
• Variable Length (16bit, 32bit, 64bit, 128bit)
  The floating number is shortened to the smallest size. Fach four integers, one byte is added to indicate the size of each integer type.

String

Only UTF-8 string is allowed, in particular, the non-UTF value of 0xFF is used internally and should be avoided.

Scalar

• Single
  Just a single Vec[u8] data

Array

• Joined
  The strings are joined with marker 0xFF, which never appears in UTF-8.
• Dictionary
  Create a dictionary of maximum 256 variants and the array is converted into an array of the references to the dictionary.

Q&A

When is it useful?

A simple case is to store a 2D array of numbers similar to a CSV file. Unlike a CSV file, ArrayObject can store arrays of more than two-dimensions, has a smaller filesize, and is fast to read and write. Instead, it is not possible to append data to ArrayObject or to have different types of data in a single ArrayObject. The definition of ArrayObject is a chunk of data that is not separable. Hence there is no sense in appending data or having different types. In other words, ArrayObject is the data that we want to load into memory simultaneously, not line by line. Another use is an object for object storage, which is actually what the ArrayObject is for. The object storage does the things that ArrayObject cannot: it allows you to append the data and store different types of data in a storage. This is why ArrayObject should not have these capabilities: having multiple options to do the same thing would complicate the system. Instead, with a minimal footer, it does what the object storage cannot: type abstraction, forced type checking and type-dependent compression.

What is the difference from a raw binary file?

A raw binary file has no type checking system. Integers, for example, can be unsigned or signed, 32bit or 64bit, little endian or big endian, etc. Without type checking, one has to check that both reader and writer are using the same format. When it comes to an array, it becomes more complicated: the raw memory order can be row major or column major depending on the programing language, and the real part and the imaginary part of complex numbers can be assigned to the most inner or the most outer index. ArrayObject has a well defined spec and one does not need to worry about these things.

What is the difference from database like HDF5?

ArrayObject is a simple, compact and portable data format for a single array, but is not a database. Thus, it does not have complicated structure like dataset or group. Even so, ArrayObject is self-contained: it knows how to inflate itself and we do not need to feed any information to read the data. Instead, ArrayObject does not have information of datasize, name, timestamp, permission, etc, which are supposed to be managed by object storage or filesystem. ArrayObject is rather closer to a CSV file: the program knows how to read it, but does not know what the data is about, when it is created, or what are the column and row, without accessing the metadata stored outside of the file.

What is the difference from serialization like Serde?

ArrayObject forbids nested structures: it can be an array of numbers or strings, but not of ArrayObjects themselves. It also cannot be a tuple. Such structures are supposed to be provided by the storage. Serialization typically adds the datasize to each data to indicate the boundaries of the data. This information is not necessary for storing data in a file or a object storage because it is already managed by the storage. Instead, ArrayObject stores the information about the shape of the array. A more technical difference is that ArrayObject adds a footer instead of a header, allowing metadata to be separated at no cost.

Why Complex Numbers?

The relation between complex numbers and real numbers is the same as that of real numbers and integers. A subset of complex numbers is real numbers and there is a well-defined map betwen them where they overlap. This makes a difference between a complex number and a vague array of length 2. Practically, in dynamically typed languages, the results of functions like sqrt or log yield either real numbers or complex numbers depending on the sign of the argument. It is thus useful to manifestly indicate complex numbers as a type and keep the same array shape. In addition, it is somewhat cumbersome to convert an array of complex numbers into an array of real numbers. ArrayObject provides a handy export/import option for complex numbers.

Why is there no conversion from Vec<Vec<_>>?

Vec<Vec<_>> may contain vectors of different lengths and does not fit into ArrayObject. In general, if you work with an fixed length array, it is much efficient to use a crate like ndarray or nalgebra. If necessary, for ArrayObjects having the same size, shape and type, .try_concat() method is available for Vec<ArrayObject>, which generates a one-dimensional higher array. On the other hand, if the lengths are different, it can be stored as multiple ArrayObjects in an object storage. See dbgbb for such an implementation.