Child pages
  • LanguageManual ORC
Skip to end of metadata
Go to start of metadata

ORC Files

ORC File Format

Version

Introduced in Hive version 0.11.

The Optimized Row Columnar (ORC) file format provides a highly efficient way to store Hive data. It was designed to overcome limitations of the other Hive file formats. Using ORC files improves performance when Hive is reading, writing, and processing data.

Compared with RCFile format, for example, ORC file format has many advantages such as:

  • a single file as the output of each task, which reduces the NameNode's load
  • Hive type support including datetime, decimal, and the complex types (struct, list, map, and union)
  • light-weight indexes stored within the file
    • skip row groups that don't pass predicate filtering
    • seek to a given row
  • block-mode compression based on data type
    • run-length encoding for integer columns
    • dictionary encoding for string columns
  • concurrent reads of the same file using separate RecordReaders
  • ability to split files without scanning for markers
  • bound the amount of memory needed for reading or writing
  • metadata stored using Protocol Buffers, which allows addition and removal of fields

File Structure

An ORC file contains groups of row data called stripes, along with auxiliary information in a file footer. At the end of the file a postscript holds compression parameters and the size of the compressed footer.

The default stripe size is 250 MB. Large stripe sizes enable large, efficient reads from HDFS.

The file footer contains a list of stripes in the file, the number of rows per stripe, and each column's data type. It also contains column-level aggregates count, min, max, and sum.

This diagram illustrates the ORC file structure:

Stripe Structure

As shown in the diagram, each stripe in an ORC file holds index data, row data, and a stripe footer.

The stripe footer contains a directory of stream locations. Row data is used in table scans.

Index data includes min and max values for each column and the row positions within each column. (A bit field or bloom filter could also be included.) Row index entries provide offsets that enable seeking to the right compression block and byte within a decompressed block.  Note that ORC indexes are used only for the selection of stripes and row groups and not for answering queries.

Having relatively frequent row index entries enables row-skipping within a stripe for rapid reads, despite large stripe sizes. By default every 10,000 rows can be skipped.

With the ability to skip large sets of rows based on filter predicates, you can sort a table on its secondary keys to achieve a big reduction in execution time. For example, if the primary partition is transaction date, the table can be sorted on state, zip code, and last name. Then looking for records in one state will skip the records of all other states.

A complete specification of the format is given in the ORC specification.

HiveQL Syntax

File formats are specified at the table (or partition) level. You can specify the ORC file format with HiveQL statements such as these:

  • CREATE TABLE ... STORED AS ORC
  • ALTER TABLE ... [PARTITION partition_spec] SET FILEFORMAT ORC
  • SET hive.default.fileformat=Orc

The parameters are all placed in the TBLPROPERTIES (see Create Table). They are:

Key

Default

Notes

orc.compress

ZLIB

high level compression (one of NONE, ZLIB, SNAPPY)

orc.compress.size

262,144

number of bytes in each compression chunk

orc.stripe.size

268435456

number of bytes in each stripe

orc.row.index.stride

10,000

number of rows between index entries (must be >= 1000)

orc.create.index

true

whether to create row indexes

orc.bloom.filter.columns""comma separated list of column names for which bloom filter should be created
orc.bloom.filter.fpp0.05false positive probability for bloom filter (must >0.0 and <1.0)

For example, creating an ORC stored table without compression:

create table Addresses (
  name string,
  street string,
  city string,
  state string,
  zip int
) stored as orc tblproperties ("orc.compress"="NONE");

Version 0.14.0+: CONCATENATE

ALTER TABLE table_name [PARTITION partition_spec] CONCATENATE can be used to merge small ORC files into a larger file, starting in Hive 0.14.0. The merge happens at the stripe level, which avoids decompressing and decoding the data.

Serialization and Compression

The serialization of column data in an ORC file depends on whether the data type is integer or string.

Integer Column Serialization

Integer columns are serialized in two streams.

  1. present bit stream: is the value non-null?
  2. data stream: a stream of integers

Integer data is serialized in a way that takes advantage of the common distribution of numbers:

  • Integers are encoded using a variable-width encoding that has fewer bytes for small integers.
  • Repeated values are run-length encoded.
  • Values that differ by a constant in the range (-128 to 127) are run-length encoded.

The variable-width encoding is based on Google's protocol buffers and uses the high bit to represent whether this byte is not the last and the lower 7 bits to encode data. To encode negative numbers, a zigzag encoding is used where 0, -1, 1, -2, and 2 map into 0, 1, 2, 3, 4, and 5 respectively.

Each set of numbers is encoded this way:

  • If the first byte (b0) is negative:
    • -b0 variable-length integers follow.
  • If the first byte (b0) is positive:
    • it represents b0 + 3 repeated integers
    • the second byte (-128 to +127) is added between each repetition
    • 1 variable-length integer.

In run-length encoding, the first byte specifies run length and whether the values are literals or duplicates. Duplicates can step by -128 to +128. Run-length encoding uses protobuf style variable-length integers.

String Column Serialization

Serialization of string columns uses a dictionary to form unique column values. The dictionary is sorted to speed up predicate filtering and improve compression ratios.

String columns are serialized in four streams.

  1. present bit stream: is the value non-null?
  2. dictionary data: the bytes for the strings
  3. dictionary length: the length of each entry
  4. row data: the row values

Both the dictionary length and the row values are run-length encoded streams of integers.

Compression

Streams are compressed using a codec, which is specified as a table property for all streams in that table. To optimize memory use, compression is done incrementally as each block is produced. Compressed blocks can be jumped over without first having to be decompressed for scanning. Positions in the stream are represented by a block start location and an offset into the block.

The codec can be Snappy, Zlib, or none.

ORC File Dump Utility

The ORC file dump utility analyzes ORC files.  To invoke it, use this command:

Adding -d to the command will cause it to dump the data in the ORC file rather than the metadata (Hive 1.1.0 and later).

Adding --rowindex with a comma separated list of column ids will cause it to print row indexes for the specified columns, where 0 is the top level struct containing all of the columns and 1 is the first column id (Hive 1.1.0 and later).

Adding -t to the command will print the timezone id of the writer.

Adding -j to the command will print the ORC file metadata in JSON format. To pretty print the JSON metadata add -p to the command.

<location-of-orc-file> is the URI of the ORC file. From Hive 1.3.0 and later, this URI can be a directory containing ORC files.

ORC Configuration Parameters

The ORC configuration parameters are described in Hive Configuration Properties – ORC File Format.

ORC Format Specification

Introduction

 

Hive's RCFile was the standard format for storing tabular data in
Hadoop for several years. However, RCFile has limitations because it
treats each column as a binary blob without semantics. In Hive 0.11 we
added a new file format named Optimized Row Columnar (ORC) file that
uses and retains the type information from the table definition. ORC
uses type specific readers and writers that provide light weight
compression techniques such as dictionary encoding, bit packing, delta
encoding, and run length encoding -- resulting in dramatically smaller
files. Additionally, ORC can apply generic compression using zlib, or
Snappy on top of the lightweight compression for even smaller
files. However, storage savings are only part of the gain. ORC
supports projection, which selects subsets of the columns for reading,
so that queries reading only one column read only the required
bytes. Furthermore, ORC files include light weight indexes that
include the minimum and maximum values for each column in each set of
10,000 rows and the entire file. Using pushdown filters from Hive, the
file reader can skip entire sets of rows that aren't important for
this query.

File Tail

Since HDFS does not support changing the data in a file after it is written, ORC stores the top level index at the end of the file. The overall structure of the file is given in the figure above.
The file's tail consists of 3 parts; the file metadata, file footer and postscript.

The metadata for ORC is stored using Protocol Buffers, which provides the ability to add new fields without breaking readers. This document incorporates the Protobuf definition from the ORC source code and the reader are encouraged to review the Protobuf encoding if they need to understand the byte-level encoding

Postscript

The Postscript section provides the necessary information to interpret
the rest of the file including the length of the file's Footer and
Metadata sections, the version of the file, and the kind of general
compression used (eg. none, zlib, or snappy). The Postscript is never
compressed and ends one byte before the end of the file. The version
stored in the Postscript is the lowest version of Hive that is
guaranteed to be able to read the file and it stored as a sequence of
the major and minor version. There are currently two versions that are
used: [0,11] for Hive 0.11, and [0,12] for Hive 0.12 to 0.14.

The process of reading an ORC file works backwards through the
file. Rather than making multiple short reads, the ORC reader reads
the last 16k bytes of the file with the hope that it will contain both
the Footer and Postscript sections. The final byte of the file
contains the serialized length of the Postscript, which must be less
than 256 bytes. Once the Postscript is parsed, the compressed
serialized length of the Footer is known and it can be decompressed
and parsed.

message PostScript {
// the length of the footer section in bytes
optional uint64 footerLength = 1;
// the kind of generic compression used
optional CompressionKind compression = 2;
// the maximum size of each compression chunk
optional uint64 compressionBlockSize = 3;
// the version of the writer
repeated uint32 version = 4 [packed = true];
// the length of the metadata section in bytes
optional uint64 metadataLength = 5;
// the fixed string "ORC"
optional string magic = 8000;
}
enum CompressionKind {
NONE = 0;
ZLIB = 1;
SNAPPY = 2;
LZO = 3;
}

The Footer section contains the layout of the body of the file, the
type schema information, the number of rows, and the statistics about
each of the columns.

The file is broken in to three parts- Header, Body, and Tail. The
Header consists of the bytes "ORC'' to support tools that want to
scan the front of the file to determine the type of the file. The Body
contains the rows and indexes, and the Tail gives the file level
information as described in this section.

message Footer {
// the length of the file header in bytes (always 3)
optional uint64 headerLength = 1;
// the length of the file header and body in bytes
optional uint64 contentLength = 2;
// the information about the stripes
repeated StripeInformation stripes = 3;
// the schema information
repeated Type types = 4;
// the user metadata that was added
repeated UserMetadataItem metadata = 5;
// the total number of rows in the file
optional uint64 numberOfRows = 6;
// the statistics of each column across the file
repeated ColumnStatistics statistics = 7;
// the maximum number of rows in each index entry
optional uint32 rowIndexStride = 8;
}

Stripe Information

The body of the file is divided into stripes. Each stripe is self
contained and may be read using only its own bytes combined with the
file's Footer and Postscript. Each stripe contains only entire rows so
that rows never straddle stripe boundaries. Stripes have three
sections: a set of indexes for the rows within the stripe, the data
itself, and a stripe footer. Both the indexes and the data sections
are divided by columns so that only the data for the required columns
needs to be read.

message StripeInformation {
// the start of the stripe within the file
optional uint64 offset = 1;
// the length of the indexes in bytes
optional uint64 indexLength = 2;
// the length of the data in bytes
optional uint64 dataLength = 3;
// the length of the footer in bytes
optional uint64 footerLength = 4;
// the number of rows in the stripe
optional uint64 numberOfRows = 5;
}

Type Information

All of the rows in an ORC file must have the same schema. Logically
the schema is expressed as a tree as in the figure below, where
the compound types have subcolumns under them.

The equivalent Hive DDL would be:

create table Foobar (
myInt int,
myMap map<string,
struct<myString : string,
myDouble: double>>,
myTime timestamp
);

The type tree is flattened in to a list via a pre-order traversal
where each type is assigned the next id. Clearly the root of the type
tree is always type id 0. Compound types have a field named subtypes
that contains the list of their children's type ids.

message Type {
enum Kind {
BOOLEAN = 0;
BYTE = 1;
SHORT = 2;
INT = 3;
LONG = 4;
FLOAT = 5;
DOUBLE = 6;
STRING = 7;
BINARY = 8;
TIMESTAMP = 9;
LIST = 10;
MAP = 11;
STRUCT = 12;
UNION = 13;
DECIMAL = 14;
DATE = 15;
VARCHAR = 16;
CHAR = 17;
}
// the kind of this type
required Kind kind = 1;
// the type ids of any subcolumns for list, map, struct, or union
repeated uint32 subtypes = 2 [packed=true];
// the list of field names for struct
repeated string fieldNames = 3;
// the maximum length of the type for varchar or char
optional uint32 maximumLength = 4;
// the precision and scale for decimal
optional uint32 precision = 5;
optional uint32 scale = 6;
}

Column Statistics

The goal of the column statistics is that for each column, the writer
records the count and depending on the type other useful fields. For
most of the primitive types, it records the minimum and maximum
values; and for numeric types it additionally stores the sum.
From Hive 1.1.0 onwards, the column statistics will also record if
there are any null values within the row group by setting the hasNull flag.
The hasNull flag is used by ORC's predicate pushdown to better answer
'IS NULL' queries.

message ColumnStatistics {
// the number of values
optional uint64 numberOfValues = 1;
// At most one of these has a value for any column
optional IntegerStatistics intStatistics = 2;
optional DoubleStatistics doubleStatistics = 3;
optional StringStatistics stringStatistics = 4;
optional BucketStatistics bucketStatistics = 5;
optional DecimalStatistics decimalStatistics = 6;
optional DateStatistics dateStatistics = 7;
optional BinaryStatistics binaryStatistics = 8;
optional TimestampStatistics timestampStatistics = 9;
optional bool hasNull = 10;
}

For integer types (tinyint, smallint, int, bigint), the column
statistics includes the minimum, maximum, and sum. If the sum
overflows long at any point during the calculation, no sum is
recorded.

message IntegerStatistics {
optional sint64 minimum = 1;
optional sint64 maximum = 2;
optional sint64 sum = 3;
}

For floating point types (float, double), the column statistics include the minimum, maximum, and sum. If the sum overflows a double, no sum is recorded.

message DoubleStatistics {
optional double minimum = 1;
optional double maximum = 2;
optional double sum = 3;
}

For strings, the minimum value, maximum value, and the sum of the lengths of the values are recorded.

message StringStatistics {
optional string minimum = 1;
optional string maximum = 2;
// sum will store the total length of all strings
optional sint64 sum = 3;
}

For booleans, the statistics include the count of false and true values.

message BucketStatistics {
repeated uint64 count = 1 [packed=true];
}

For decimals, the minimum, maximum, and sum are stored.

message DecimalStatistics {
optional string minimum = 1;
optional string maximum = 2;
optional string sum = 3;
}

Date columns record the minimum and maximum values as the number of days since
the epoch (1/1/2015).

message DateStatistics {
// min,max values saved as days since epoch
optional sint32 minimum = 1;
optional sint32 maximum = 2;
}

Timestamp columns record the minimum and maximum values as the number of
milliseconds since the epoch (1/1/2015).

message TimestampStatistics {
// min,max values saved as milliseconds since epoch
optional sint64 minimum = 1;
optional sint64 maximum = 2;
}

Binary columns store the aggregate number of bytes across all of the values.

message BinaryStatistics {
// sum will store the total binary blob length
optional sint64 sum = 1;
}

User Metadata

The user can add arbitrary key/value pairs to an ORC file as it is
written. The contents of the keys and values are completely
application defined, but the key is a string and the value is
binary. Care should be taken by applications to make sure that their
keys are unique and in general should be prefixed with an organization
code.

message UserMetadataItem {
// the user defined key
required string name = 1;
// the user defined binary value
required bytes value = 2;
}

File Metadata

The file Metadata section contains column statistics at the stripe
level granularity. These statistics enable input split elimination
based on the predicate push-down evaluated per a stripe.

message StripeStatistics {
repeated ColumnStatistics colStats = 1;
}
message Metadata {
repeated StripeStatistics stripeStats = 1;
}

Compression Streams

If the ORC file writer selects a generic compression codec (zlib or
snappy), every part of the ORC file except for the Postscript is
compressed with that codec. However, one of the requirements for ORC
is that the reader be able to skip over compressed bytes without
decompressing the entire stream. To manage this, ORC writes compressed
streams in chunks with headers as in the figure below.
To handle uncompressable data, if the compressed data is larger than
the original, the original is stored and the isOriginal flag is
set. Each header is 3 bytes long with (compressedLength * 2 +
isOriginal) stored as a little endian value. For example, the header
for a chunk that compressed to 100,000 bytes would be [0x40, 0x0d,
0x03]. The header for 5 bytes that did not compress would be [0x0b,
0x00, 0x00]. Each compression chunk is compressed independently so
that as long as a decompressor starts at the top of a header, it can
start decompressing without the previous bytes.

The default compression chunk size is 256K, but writers can choose
their own value less than 223. Larger chunks lead to better
compression, but require more memory. The chunk size is recorded in
the Postscript so that readers can allocate appropriately sized
buffers.

ORC files without generic compression write each stream directly
with no headers.

Run Length Encoding

Base 128 Varint

Variable width integer encodings take advantage of the fact that most
numbers are small and that having smaller encodings for small numbers
shrinks the overall size of the data. ORC uses the varint format from
Protocol Buffers, which writes data in little endian format using the
low 7 bits of each byte. The high bit in each byte is set if the
number continues into the next byte.

Unsigned OriginalSerialized
00x00
10x01
1270x7f
1280x80, 0x01
1290x81, 0x01
163830xff, 0x7f
163840x80, 0x80, 0x01
163850x81, 0x80, 0x01

For signed integer types, the number is converted into an unsigned
number using a zigzag encoding. Zigzag encoding moves the sign bit to
the least significant bit using the expression (val << 1) ^ (val >> 63) and derives its name from the fact that positive and negative
numbers alternate once encoded. The unsigned number is then serialized
as above.

Signed OriginalUnsigned
00
-11
12
-23
24

Byte Run Length Encoding

For byte streams, ORC uses a very light weight encoding of identical
values.

  • Run - a sequence of at least 3 identical values
  • Literals - a sequence of non-identical values

The first byte of each group of values is a header than determines
whether it is a run (value between 0 to 127) or literal list (value
between -128 to -1). For runs, the control byte is the length of the
run minus the length of the minimal run (3) and the control byte for
literal lists is the negative length of the list. For example, a
hundred 0's is encoded as [0x61, 0x00] and the sequence 0x44, 0x45
would be encoded as [0xfe, 0x44, 0x45]. The next group can choose
either of the encodings.

Boolean Run Length Encoding

For encoding boolean types, the bits are put in the bytes from most
significant to least significant. The bytes are encoded using byte run
length encoding as described in the previous section. For example,
the byte sequence [0xff, 0x80] would be one true followed by
seven false values.

Integer Run Length Encoding version 1

In Hive 0.11 ORC files used Run Length Encoding version 1 (RLEv1),
which provides a lightweight compression of signed or unsigned integer
sequences. RLEv1 has two sub-encodings:

  • Run - a sequence of values that differ by a small fixed delta
  • Literals - a sequence of varint encoded values

Runs start with an initial byte of 0x00 to 0x7f, which encodes the
length of the run - 3. A second byte provides the fixed delta in the
range of -128 to 127. Finally, the first value of the run is encoded
as a base 128 varint.

For example, if the sequence is 100 instances of 7 the encoding would
start with 100 - 3, followed by a delta of 0, and a varint of 7 for
an encoding of [0x61, 0x00, 0x07]. To encode the sequence of numbers
running from 100 to 1, the first byte is 100 - 3, the delta is -1,
and the varint is 100 for an encoding of [0x61, 0xff, 0x64].

Literals start with an initial byte of 0x80 to 0xff, which corresponds
to the negative of number of literals in the sequence. Following the
header byte, the list of N varints is encoded. Thus, if there are
no runs, the overhead is 1 byte for each 128 integers. The first 5
prime numbers [2, 3, 4, 7, 11] would encoded as [0xfb, 0x02, 0x03,
0x04, 0x07, 0xb].

Integer Run Length Encoding Version 2

In Hive 0.12, ORC introduced Run Length Encoding version 2 (RLEv2),
which has improved compression and fixed bit width encodings for
faster expansion. RLEv2 uses four sub-encodings based on the data:

  • Short Repeat - used for short sequences with repeated values
  • Direct - used for random sequences with a fixed bit width
  • Patched Base - used for random sequences with a variable bit width
  • Delta - used for monotonically increasing or decreasing sequences

Short Repeat

The short repeat encoding is used for short repeating integer
sequences with the goal of minimizing the overhead of the header. All
of the bits listed in the header are from the first byte to the last
and from most significant bit to least significant bit. If the type is
signed, the value is zigzag encoded.

  • 1 byte header
    • 2 bits for encoding type (0)
    • 3 bits for width (W) of repeating value (1 to 8 bytes)
    • 3 bits for repeat count (3 to 10 values)
  • W bytes in big endian format, which is zigzag encoded if they type is signed

The unsigned sequence of [10000, 10000, 10000, 10000, 10000] would be serialized with short repeat encoding (0), a width of 2 bytes (1), and repeat count of 5 (2) as [0x0a, 0x27, 0x10].

Direct

The direct encoding is used for integer sequences whose values have a
relatively constant bit width. It encodes the values directly using a
fixed width big endian encoding. The width of the values is encoded
using the table below.

 

The 5 bit width encoding table for RLEv2
Width in BitsEncoded ValueNotes
00for delta encoding
10for non-delta encoding
21 
43 
87 
1615 
2423 
3227 
4028 
4829 
5630 
6431 
32deprecated
5 <= x <= 7x - 1deprecated
9 <= x <= 15x -1deprecated
17 <= x <= 21x - 1deprecated
2624deprecated
2825deprecated
3026deprecated
  • 2 bytes header
    • 2 bits for encoding type (1)
    • 5 bits for encoded width (W) of values (1 to 64 bits) using the 5 bit width encoding table
    • 9 bits for length (L) (1 to 512 values)
  • W * L bits (padded to the next byte) encoded in big endian format, which is
    zigzag encoding if the type is signed

The unsigned sequence of [23713, 43806, 57005, 48879] would be
serialized with direct encoding (1), a width of 16 bits (15), and
length of 4 (3) as [0x5e, 0x03, 0x5c, 0xa1, 0xab, 0x1e, 0xde, 0xad,
0xbe, 0xef].

Patched Base

The patched base encoding is used for integer sequences whose bit
widths varies a lot. The minimum signed value of the sequence is found
and subtracted from the other values. The bit width of those adjusted
values is analyzed and the 90 percentile of the bit width is chosen
as W. The 10\% of values larger than W use patches from a patch list
to set the additional bits. Patches are encoded as a list of gaps in
the index values and the additional value bits.

  • 4 bytes header
    • bits for encoding type (2)
    • 5 bits for encoded width (W) of values (1 to 64 bits)
      using the 5 bit width encoding table
    • 9 bits for length (L) (1 to 512 values)
    • 3 bits for base value width (BW) (1 to 8 bytes)
    • 5 bits for patch width (PW) (1 to 64 bits)
      using  the 5 bit width encoding table
    • 3 bits for patch gap width (PGW) (1 to 8 bits)
    • 5 bits for patch list length (PLL) (0 to 31 patches)
  • Base value (BW bytes) - The base value is stored as a big endian value
    with negative values marked by the most significant bit set. If it that
    bit is set, the entire value is negated.
  • Data values (W * L bits padded to the byte) - A sequence of W bit positive values that
    are added to the base value.
  • Data values (W * L bits padded to the byte) - A sequence of W bit positive values that are added to the base value.
  • Patch list (PLL * (PGW + PW) bytes) - A list of patches for values that didn't fit within W bits. Each entry in the list consists of a gap, which is the number of elements skipped from the previous patch, and a patch value. Patches are applied by logically or'ing the data values with the relevant patch shifted W bits left. If a patch is 0, it was introduced to skip over more than 255 items. The combined length of each patch (PGW + PW) must be less or equal to 64.

The unsigned sequence of [2030, 2000, 2020, 1000000, 2040, 2050, 2060,
2070, 2080, 2090] has a minimum of 2000, which makes the adjusted
sequence [30, 0, 20, 998000, 40, 50, 60, 70, 80, 90]. It has an
encoding of patched base (2), a bit width of 8 (7), a length of 10
(9), a base value width of 2 bytes (1), a patch width of 12 bits (11),
patch gap width of 2 bits (1), and a patch list length of 1 (1). The
base value is 2000 and the combined result is [0x8e, 0x09, 0x2b, 0x21,
0x07, 0xd0, 0x1e, 0x00, 0x14, 0x70, 0x28, 0x32, 0x3c, 0x46, 0x50,
0x5a, 0xfc, 0xe8]

Delta

The Delta encoding is used for monotonically increasing or decreasing
sequences. The first two numbers in the sequence can not be identical,
because the encoding is using the sign of the first delta to determine
if the series is increasing or decreasing.

  • 2 bytes header
    • 2 bits for encoding type (3)
    • 5 bits for encoded width (W) of deltas (0 to 64 bits)
      using the 5 bit width encoding table
    • 9 bits for run length (L) (1 to 512 values)
  • Base value - encoded as (signed or unsigned) varint
  • Delta base - encoded as signed varint
  • Delta values $W * (L - 2)$ bytes - encode each delta after the first
    one. If the delta base is positive, the sequence is increasing and if it is
    negative the sequence is decreasing.

The unsigned sequence of [2, 3, 5, 7, 11, 13, 17, 19, 23, 29] would be serialized with delta encoding (3), a width of 4 bits (3), length of 10 (9), a base of 2 (2), and first delta of 1 (2). The resulting sequence is [0xc6, 0x09, 0x02, 0x02, 0x22, 0x42, 0x42, 0x46].

Stripes

The body of ORC files consists of a series of stripes. Stripes are
large (typically ~200MB) and independent of each other and are often
processed by different tasks. The defining characteristic for columnar
storage formats is that the data for each column is stored separately
and that reading data out of the file should be proportional to the
number of columns read.

In ORC files, each column is stored in several streams that are stored
next to each other in the file. For example, an integer column is
represented as two streams PRESENT, which uses one with a bit per
value recording if the value is non-null, and DATA, which records the
non-null values. If all of a column's values in a stripe are non-null,
the PRESENT stream is omitted from the stripe. For binary data, ORC
uses three streams PRESENT, DATA, and LENGTH, which stores the length
of each value. The details of each type will be presented in the
following subsections.

Stripe Footer

The stripe footer contains the encoding of each column and the
directory of the streams including their location.

message StripeFooter {
// the location of each stream
repeated Stream streams = 1;
// the encoding of each column
repeated ColumnEncoding columns = 2;
}

To describe each stream, ORC stores the kind of stream, the column id,
and the stream's size in bytes. The details of what is stored in each stream
depends on the type and encoding of the column.

message Stream {
enum Kind {
// boolean stream of whether the next value is non-null
PRESENT = 0;
// the primary data stream
DATA = 1;
// the length of each value for variable length data
LENGTH = 2;
// the dictionary blob
DICTIONARY\_DATA = 3;
// deprecated prior to Hive 0.11
// It was used to store the number of instances of each value in the
// dictionary
DICTIONARY_COUNT = 4;
// a secondary data stream
SECONDARY = 5;
// the index for seeking to particular row groups
ROW_INDEX = 6;
}
required Kind kind = 1;
// the column id
optional uint32 column = 2;
// the number of bytes in the file
optional uint64 length = 3;
}

Depending on their type several options for encoding are possible. The
encodings are divided into direct or dictionary-based categories and
further refined as to whether they use RLE v1 or v2.

message ColumnEncoding {
enum Kind {
// the encoding is mapped directly to the stream using RLE v1
DIRECT = 0;
// the encoding uses a dictionary of unique values using RLE v1
DICTIONARY = 1;
// the encoding is direct using RLE v2
DIRECT\_V2 = 2;
// the encoding is dictionary-based using RLE v2
DICTIONARY\_V2 = 3;
}
required Kind kind = 1;
// for dictionary encodings, record the size of the dictionary
optional uint32 dictionarySize = 2;
}

Column Encodings

SmallInt, Int, and BigInt Columns

All of the 16, 32, and 64 bit integer column types use the same set of
potential encodings, which is basically whether they use RLE v1 or
v2. If the PRESENT stream is not included, all of the values are
present. For values that have false bits in the present stream, no
values are included in the data stream.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoSigned Integer RLE v1
DIRECT_V2PRESENTYes

Boolean RLE

 DATANoSigned Integer RLE v2

Float and Double Columns

Floating point types are stored using IEEE 754 floating point bit
layout. Float columns use 4 bytes per value and double columns use 8
bytes.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoIEEE 754 floating point representation

String, Char, and VarChar Columns

String columns are adaptively encoded based on whether the first
10,000 values are sufficiently distinct. In all of the encodings, the
PRESENT stream encodes whether the value is null.

For direct encoding the UTF-8 bytes are saved in the DATA stream and
the length of each value is written into the LENGTH stream. In direct
encoding, if the values were [``Nevada'', ``California'']; the DATA
would be ``NevadaCalifornia'' and the LENGTH would be [6, 10].

For dictionary encodings the dictionary is sorted and UTF-8 bytes of
each unique value are placed into DICTIONARY_DATA. The length of each
item in the dictionary is put into the LENGTH stream. The DATA stream
consists of the sequence of references to the dictionary elements.

In dictionary encoding, if the values were [``Nevada'',
``California'', ``Nevada'', ``California'', and ``Florida'']; the
DICTIONARY\_DATA would be ``CaliforniaFloridaNevada'' and LENGTH would
be [10, 7, 6]. The DATA would be [2, 0, 2, 0, 1].

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoString contents
 LENGTHNoUnsigned Integer RLE v1
DICTIONARYPRESENTYesBoolean RLE
 DATANoUnsigned Integer RLE v1
 DICTIONARY_DATANoString contents
 LENGTHNoUnsigned Integer RLE v1
DIRECT_V2PRESENTYesBoolean RLE
 DATANoString Contents
 LENGTHNoUnsigned Integer RLE v2
DICTIONARY_V2PRESENTYesBoolean RLE
 DATANoUnsigned Integer RLE v2
 DICTIONARY_DATANoString contents
 LENGTHNoUnsigned Integer RLE v2

Boolean Columns

Boolean columns are rare, but have a simple encoding.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoBoolean RLE

TinyInt Columns

TinyInt (byte) columns use byte run length encoding.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoByte RLE

Binary Columns

Binary data is encoded with a PRESENT stream, a DATA stream that records
the contents, and a LENGTH stream that records the number of bytes per a
value.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoBinary contents
 LENGTHNoUnsigned Integer RLE v1
DIRECT_V2PRESENTYesBoolean RLE
 DATANoBinary contents
 LENGTHNoUnsigned Integer RLE v2

Decimal Columns

Decimal was introduced in Hive 0.11 with infinite precision (the total
number of digits). In Hive 0.13, the definition was change to limit
the precision to a maximum of 38 digits, which conveniently uses 127
bits plus a sign bit. The current encoding of decimal columns stores
the integer representation of the value as an unbounded length zigzag
encoded base 128 varint. The scale is stored in the SECONDARY stream
as an unsigned integer.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoUnbounded base 128 varints
 SECONDARYNoUnsigned Integer RLE v1
DIRECT_V2PRESENTYesBoolean RLE
 DATANoUnbounded base 128 varints
 SECONDARYNoUnsigned Integer RLE v2

Date Columns

Date data is encoded with a PRESENT stream, a DATA stream that records
the number of days after January 1, 1970 in UTC.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoSigned Integer RLE v1
DIRECT_V2PRESENTYesBoolean RLE
 DATANoSigned Integer RLE v2

Timestamp Columns

Timestamp records times down to nanoseconds as a PRESENT stream that
records non-null values, a DATA stream that records the number of
seconds after 1 January 2015, and a SECONDARY stream that records the
number of nanoseconds.

Because the number of nanoseconds often has a large number of trailing
zeros, the number has trailing decimal zero digits removed and the
last three bits are used to record how many zeros were removed. Thus
1000 nanoseconds would be serialized as 0x0b and 100000 would be
serialized as 0x0d.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoSigned Integer RLE v1
 SECONDARYNoUnsigned Integer RLE v1
DIRECT_V2PRESENTYesBoolean RLE
 DATANoSigned Integer RLE v2
 SECONDARYNoUnsigned Integer RLE v2

Struct Columns

Structs have no data themselves and delegate everything to their child
columns except for their PRESENT stream. They have a child column
for each of the fields.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE

List Columns

Lists are encoded as the PRESENT stream and a length stream with
number of items in each list. They have a single child column for the
element values.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 LENGTHNoUnsigned Integer RLE v1
DIRECT_V2PRESENTYesBoolean RLE
 LENGTHNoUnsigned Integer RLE v2

Map Columns

Maps are encoded as the PRESENT stream and a length stream with number
of items in each list. They have a child column for the key and
another child column for the value.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 LENGTHNoUnsigned Integer RLE v1
DIRECT_V2PRESENTYesBoolean RLE
 LENGTHNoUnsigned Integer RLE v2

Union Columns

Unions are encoded as the PRESENT stream and a tag stream that controls which
potential variant is used. They have a child column for each variant of the
union. Currently ORC union types are limited to 256 variants, which matches
the Hive type model.

EncodingStream KindOptionalContents
DIRECTPRESENTYesBoolean RLE
 DATANoByte RLE

Indexes

Row Group Index

The row group indexes consist of a ROW\_INDEX stream for each primitive
column that has an entry for each row group. Row groups are controlled
by the writer and default to 10,000 rows. Each RowIndexEntry gives the
position of each stream for the column and the statistics for that row
group.

The index streams are placed at the front of the stripe, because in
the default case of streaming they do not need to be read. They are
only loaded when either predicate push down is being used or the
reader seeks to a particular row.

message RowIndexEntry {
repeated uint64 positions = 1 [packed=true];
optional ColumnStatistics statistics = 2;
}
message RowIndex {
repeated RowIndexEntry entry = 1;
}

To record positions, each stream needs a sequence of numbers. For
uncompressed streams, the position is the byte offset of the RLE run's
start location followed by the number of values that need to be
consumed from the run. In compressed streams, the first number is the
start of the compression chunk in the stream, followed by the number
of decompressed bytes that need to be consumed, and finally the number
of values consumed in the RLE.

For columns with multiple streams, the sequences of positions in each
stream are concatenated. That was an unfortunate decision on my part
that we should fix at some point, because it makes code that uses the
indexes error-prone.

Because dictionaries are accessed randomly, there is not a position to
record for the dictionary and the entire dictionary must be read even
if only part of a stripe is being read.

Bloom Filter Index

Version 1.2.0+: Bloom Filter

Bloom Filters are added to ORC indexes from Hive 1.2.0 onwards.
Predicate pushdown can make use of bloom filters to better prune
the row groups that do not satisfy the filter condition.

The bloom filter indexes consist of a BLOOM_FILTER stream for each
column specified through 'orc.bloom.filter.columns' table properties.
A BLOOM_FILTER stream records a bloom filter entry for each row
group (default to 10,000 rows) in a column. Only the row groups that
satisfy min/max row index evaluation will be evaluated against the
bloom filter index.

Each BloomFilterEntry stores the number of hash functions ('k') used and
the bitset backing the bloom filter. The bitset is serialized as repeated
longs from which the number of bits ('m') for the bloom filter can be derived.
m = bitset.length * 64.

message BloomFilter {
optional uint32 numHashFunctions = 1;
repeated fixed64 bitset = 2;
}
message BloomFilterIndex {
repeated BloomFilter bloomFilter = 1;
}

Bloom filter internally uses two different hash functions to map a key
to a position in the bit set. For tinyint, smallint, int, bigint, float
and double types, Thomas Wang's 64-bit integer hash function is used.
Floats are converted to IEEE-754 32 bit representation
(using Java's Float.floatToIntBits(float)). Similary, Doubles are
converted to IEEE-754 64 bit representation (using Java's
Double.doubleToLongBits(double)). All these primitive types
are cast to long base type before being passed on to the hash function.
For strings and binary types, Murmur3 64 bit hash algorithm is used.
The 64 bit variant of Murmur3 considers only the most significant
8 bytes of Murmur3 128-bit algorithm. The 64 bit hashcode generated
from the above algorithms is used as a base to derive 'k' different
hash functions. We use the idea mentioned in the paper "Less Hashing,
Same Performance: Building a Better Bloom Filter" by Kirsch et. al. to
quickly compute the k hashcodes.

The algorithm for computing k hashcodes and setting the bit position
in a bloom filter is as follows:

  1. Get 64 bit base hash code from Murmur3 or Thomas Wang's hash algorithm.
  2. Split the above hashcode into two 32-bit hashcodes (say hash1 and hash2).
  3. k'th hashcode is obtained by (where k > 0):
      combinedHash = hash1 + (k * hash2)
  4. If combinedHash is negative flip all the bits:
      combinedHash = ~combinedHash
  5. Bit set position is obtained by performing modulo with m:
      position = combinedHash % m
  6. Set the position in bit set. The LSB 6 bits identifies the long index within bitset
    and bit position within the long uses little endian order.
      bitset[position >>> 6] |= (1L << position);

Bloom filter streams are interlaced with row group indexes. This placement 
makes it convenient to read the bloom filter stream and row index stream
together in single read operation.

  • No labels