Class OrderedBytes

java.lang.Object
org.apache.hadoop.hbase.util.OrderedBytes

@Public public class OrderedBytes extends Object
Utility class that handles ordered byte arrays. That is, unlike Bytes, these methods produce byte arrays which maintain the sort order of the original values.

Encoding Format summary

Each value is encoded as one or more bytes. The first byte of the encoding, its meaning, and a terse description of the bytes that follow is given by the following table:

Content Type Encoding
NULL 0x05
negative infinity 0x07
negative large 0x08, ~E, ~M
negative medium 0x13-E, ~M
negative small 0x14, -E, ~M
zero 0x15
positive small 0x16, ~-E, M
positive medium 0x17+E, M
positive large 0x22, E, M
positive infinity 0x23
NaN 0x25
fixed-length 32-bit integer 0x27, I
fixed-length 64-bit integer 0x28, I
fixed-length 8-bit integer 0x29
fixed-length 16-bit integer 0x2a
fixed-length 32-bit float 0x30, F
fixed-length 64-bit float 0x31, F
TEXT 0x33, T
variable length BLOB 0x35, B
byte-for-byte BLOB 0x36, X

Null Encoding

Each value that is a NULL encodes as a single byte of 0x05. Since every other value encoding begins with a byte greater than 0x05, this forces NULL values to sort first.

Text Encoding

Each text value begins with a single byte of 0x33 and ends with a single byte of 0x00. There are zero or more intervening bytes that encode the text value. The intervening bytes are chosen so that the encoding will sort in the desired collating order. The intervening bytes may not contain a 0x00 character; the only 0x00 byte allowed in a text encoding is the final byte.

The text encoding ends in 0x00 in order to ensure that when there are two strings where one is a prefix of the other that the shorter string will sort first.

Binary Encoding

There are two encoding strategies for binary fields, referred to as "BlobVar" and "BlobCopy". BlobVar is less efficient in both space and encoding time. It has no limitations on the range of encoded values. BlobCopy is a byte-for-byte copy of the input data followed by a termination byte. It is extremely fast to encode and decode. It carries the restriction of not allowing a 0x00 value in the input byte[] as this value is used as the termination byte.

BlobVar

"BlobVar" encodes the input byte[] in a manner similar to a variable length integer encoding. As with the other OrderedBytes encodings, the first encoded byte is used to indicate what kind of value follows. This header byte is 0x37 for BlobVar encoded values. As with the traditional varint encoding, the most significant bit of each subsequent encoded byte is used as a continuation marker. The 7 remaining bits contain the 7 most significant bits of the first unencoded byte. The next encoded byte starts with a continuation marker in the MSB. The least significant bit from the first unencoded byte follows, and the remaining 6 bits contain the 6 MSBs of the second unencoded byte. The encoding continues, encoding 7 bytes on to 8 encoded bytes. The MSB of the final encoded byte contains a termination marker rather than a continuation marker, and any remaining bits from the final input byte. Any trailing bits in the final encoded byte are zeros.

BlobCopy

"BlobCopy" is a simple byte-for-byte copy of the input data. It uses 0x38 as the header byte, and is terminated by 0x00 in the DESCENDING case. This alternative encoding is faster and more space-efficient, but it cannot accept values containing a 0x00 byte in DESCENDING order.

Variable-length Numeric Encoding

Numeric values must be coded so as to sort in numeric order. We assume that numeric values can be both integer and floating point values. Clients must be careful to use inspection methods for encoded values (such as isNumericInfinite(PositionedByteRange) and isNumericNaN(PositionedByteRange) to protect against decoding values into object which do not support these numeric concepts (such as Long and BigDecimal).

Simplest cases first: If the numeric value is a NaN, then the encoding is a single byte of 0x25. This causes NaN values to sort after every other numeric value.

If the numeric value is a negative infinity then the encoding is a single byte of 0x07. Since every other numeric value except NaN has a larger initial byte, this encoding ensures that negative infinity will sort prior to every other numeric value other than NaN.

If the numeric value is a positive infinity then the encoding is a single byte of 0x23. Every other numeric value encoding begins with a smaller byte, ensuring that positive infinity always sorts last among numeric values. 0x23 is also smaller than 0x33, the initial byte of a text value, ensuring that every numeric value sorts before every text value.

If the numeric value is exactly zero then it is encoded as a single byte of 0x15. Finite negative values will have initial bytes of 0x08 through 0x14 and finite positive values will have initial bytes of 0x16 through 0x22.

For all numeric values, we compute a mantissa M and an exponent E. The mantissa is a base-100 representation of the value. The exponent E determines where to put the decimal point.

Each centimal digit of the mantissa is stored in a byte. If the value of the centimal digit is X (hence X≥0 and X≤99) then the byte value will be 2*X+1 for every byte of the mantissa, except for the last byte which will be 2*X+0. The mantissa must be the minimum number of bytes necessary to represent the value; trailing X==0 digits are omitted. This means that the mantissa will never contain a byte with the value 0x00.

If we assume all digits of the mantissa occur to the right of the decimal point, then the exponent E is the power of one hundred by which one must multiply the mantissa to recover the original value.

Values are classified as large, medium, or small according to the value of E. If E is 11 or more, the value is large. For E between 0 and 10, the value is medium. For E less than zero, the value is small.

Large positive values are encoded as a single byte 0x22 followed by E as a varint and then M. Medium positive values are a single byte of 0x17+E followed by M. Small positive values are encoded as a single byte 0x16 followed by the ones-complement of the varint for -E followed by M.

Small negative values are encoded as a single byte 0x14 followed by -E as a varint and then the ones-complement of M. Medium negative values are encoded as a byte 0x13-E followed by the ones-complement of M. Large negative values consist of the single byte 0x08 followed by the ones-complement of the varint encoding of E followed by the ones-complement of M.

Fixed-length Integer Encoding

All 4-byte integers are serialized to a 5-byte, fixed-width, sortable byte format. All 8-byte integers are serialized to the equivelant 9-byte format. Serialization is performed by writing a header byte, inverting the integer sign bit and writing the resulting bytes to the byte array in big endian order.

Fixed-length Floating Point Encoding

32-bit and 64-bit floating point numbers are encoded to a 5-byte and 9-byte encoding format, respectively. The format is identical, save for the precision respected in each step of the operation.

This format ensures the following total ordering of floating point values: Float.NEGATIVE_INFINITY < -Float.MAX_VALUE < ... < -Float.MIN_VALUE < -0.0 < +0.0; < Float.MIN_VALUE < ... < Float.MAX_VALUE < Float.POSITIVE_INFINITY < Float.NaN

Floating point numbers are encoded as specified in IEEE 754. A 32-bit single precision float consists of a sign bit, 8-bit unsigned exponent encoded in offset-127 notation, and a 23-bit significand. The format is described further in the Single Precision Floating Point Wikipedia page

The value of a normal float is -1 sign bit × 2exponent - 127 × 1.significand

The IEE754 floating point format already preserves sort ordering for positive floating point numbers when the raw bytes are compared in most significant byte order. This is discussed further at http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm

Thus, we need only ensure that negative numbers sort in the the exact opposite order as positive numbers (so that say, negative infinity is less than negative 1), and that all negative numbers compare less than any positive number. To accomplish this, we invert the sign bit of all floating point numbers, and we also invert the exponent and significand bits if the floating point number was negative.

More specifically, we first store the floating point bits into a 32-bit int j using Float.floatToIntBits(float). This method collapses all NaNs into a single, canonical NaN value but otherwise leaves the bits unchanged. We then compute

 j ˆ= (j >> (Integer.SIZE - 1)) | Integer.MIN_SIZE
 

which inverts the sign bit and XOR's all other bits with the sign bit itself. Comparing the raw bytes of j in most significant byte order is equivalent to performing a single precision floating point comparison on the underlying bits (ignoring NaN comparisons, as NaNs don't compare equal to anything when performing floating point comparisons).

The resulting integer is then converted into a byte array by serializing the integer one byte at a time in most significant byte order. The serialized integer is prefixed by a single header byte. All serialized values are 5 bytes in length.

OrderedBytes encodings are heavily influenced by the SQLite4 Key Encoding. Slight deviations are make in the interest of order correctness and user extensibility. Fixed-width Long and Double encodings are based on implementations from the now defunct Orderly library.