tustvold commented on code in PR #264:
URL: https://github.com/apache/arrow-site/pull/264#discussion_r1013734348
##########
_posts/2022-10-30-multi-column-sorts-in-arrow-rust-part-2.md:
##########
@@ -0,0 +1,243 @@
+---
+layout: post
+title: "Fast and Memory Efficient Multi-Column Sorts in Apache Arrow Rust,
Part 2"
+date: "2022-10-30 00:00:00"
+author: "tustvold and alamb"
+categories: [arrow]
+---
+<!--
+{% comment %}
+Licensed to the Apache Software Foundation (ASF) under one or more
+contributor license agreements. See the NOTICE file distributed with
+this work for additional information regarding copyright ownership.
+The ASF licenses this file to you under the Apache License, Version 2.0
+(the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+{% endcomment %}
+-->
+
+## Introduction
+
+In Part 1 <!-- TODO add link --> of this post, we described the problem of
Multi-Column Sorting and the challenges of implementing it efficiently. This
second post explains how the new [row
format](https://docs.rs/arrow/25.0.0/arrow/row/index.html) in the [Rust
implementation](https://github.com/apache/arrow-rs) of [Apache
Arrow](https://arrow.apache.org/) works and is constructed.
+
+
+## Row Format
+
+The row format is a variable length byte sequence created by concatenating the
encoded form of each column. The encoding for each column depends on its
datatype (and sort options).
+
+```
+ ┌─────┐ ┌─────┐ ┌─────┐
+ │ │ │ │ │ │
+ ├─────┤ ┌ ┼─────┼ ─ ┼─────┼ ┐ ┏━━━━━━━━━━━━━┓
+ │ │ │ │ │ │ ─────────────▶┃ ┃
+ ├─────┤ └ ┼─────┼ ─ ┼─────┼ ┘ ┗━━━━━━━━━━━━━┛
+ │ │ │ │ │ │
+ └─────┘ └─────┘ └─────┘
+ ...
+ ┌─────┐ ┌ ┬─────┬ ─ ┬─────┬ ┐ ┏━━━━━━━━┓
+ │ │ │ │ │ │ ─────────────▶┃ ┃
+ └─────┘ └ ┴─────┴ ─ ┴─────┴ ┘ ┗━━━━━━━━┛
+ Customer State Orders
+ UInt64 Utf8 F64
+
+ Input Arrays Row Format
+ (Columns)
+```
+
+### Unsigned Integers
+
+To encode a non-null unsigned integer, the byte `0x01` is written, followed by
the integer’s bytes starting with the most significant, i.e. big endian. A null
is encoded as a `0x00` byte, followed by the encoded bytes of the integer’s
zero value
+
+```
+ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┬──┐
+ 3 │03│00│00│00│ │01│00│00│00│03│
+ └──┴──┴──┴──┘ └──┴──┴──┴──┴──┘
+ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┬──┐
+ 258 │02│01│00│00│ │01│00│00│01│02│
+ └──┴──┴──┴──┘ └──┴──┴──┴──┴──┘
+ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┬──┐
+ 23423 │7F│5B│00│00│ │01│00│00│5B│7F│
+ └──┴──┴──┴──┘ └──┴──┴──┴──┴──┘
+ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┬──┐
+ NULL │??│??│??│??│ │00│00│00│00│00│
+ └──┴──┴──┴──┘ └──┴──┴──┴──┴──┘
+
+ 32-bit (4 bytes) Row Format
+ Value Little Endian
+```
+
+### Signed Integers
+
+In Rust and most modern computer architectures, signed integers are encoded
using [two's complement](https://en.wikipedia.org/wiki/Two%27s_complement),
where a number is negated by flipping all the bits, and adding 1. Therefore,
flipping the top-most bit and treating the result as an unsigned integer
preserves the order. This unsigned integer can then be encoded using the same
encoding for unsigned integers described in the previous section. For example
+
+```
+ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┬──┐
+ 5 │05│00│00│00│ │85│00│00│00│ │01│00│00│00│85│
+ └──┴──┴──┴──┘ └──┴──┴──┴──┘ └──┴──┴──┴──┴──┘
+ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┐ ┌──┬──┬──┬──┬──┐
+ -5 │FB│FF│FF│FF│ │7B│FF│FF│FF│ │01│FF│FF│FF│FF│
+ └──┴──┴──┴──┘ └──┴──┴──┴──┘ └──┴──┴──┴──┴──┘
+
+ Value 32-bit (4 bytes) High bit flipped Row Format
+ Little Endian
+```
+
+### Floating Point
+
+Floating point values can be ordered according to the [IEEE 754 totalOrder
predicate](https://en.wikipedia.org/wiki/IEEE_754#Total-ordering_predicate)
(implemented in Rust by
[f32::total_cmp](https://doc.rust-lang.org/std/primitive.f32.html#method.total_cmp)).
This ordering interprets the bytes of the floating point value as the
correspondingly sized, signed, little-endian integer, flipping all the bits
except the sign bit in the case of negatives.
+
+Floating point values are therefore encoded to row format by converting them
to the appropriate sized signed integer representation, and then using the same
encoding for signed integers described in the previous section.
+
+### Byte Arrays (Including Strings)
+
+Unlike primitive types above, byte arrays are variable length. For short
strings, such as `state` in our example above, it is possible to pad all values
to the length of the longest one with some fixed value such as `0x00` and
produce a fixed length row. This is the approach described in the DuckDB blog
for encoding `c_birth_country`.
+
+However, often values in string columns differ substantially in length or the
maximum length is not known at the start of execution, making it inadvisable
and/or impractical to pad the strings to a fixed length. The Rust Arrow row
format therefore uses a variable length encoding.
+
+We need an encoding that unambiguously terminates the end of the byte array.
This not only permits recovering the original value from the row format, but
ensures that bytes of a longer byte array are not compared against bytes from a
different column when compared against a row containing a shorter byte array.
+
+A null byte array is encoded as a single `0x00` byte. Similarly, an empty byte
array is encoded as a single `0x01` byte.
+
+To encode a non-null, non-empty array, first a single `0x02` byte is written.
Then the array is written in 32-byte blocks, with each block followed by a
`0xFF` byte as a continuation token. The final block is padded to 32-bytes with
`0x00`, and is then followed by the unpadded length of this final block as a
single byte.
+
+Note the following example encodings use a block size of 4 bytes, as opposed
to 32 bytes for brevity
+
+```
+ ┌───┬───┬───┬───┬───┬───┬───┐
+ "MEEP" │02 │'M'│'E'│'E'│'P'│FF │04 │
+ └───┴───┴───┴───┴───┴───┴───┘
+
+ ┌───┬───┬───┬───┬───┬───┬───┐
+ "" │02 │00 │00 │00 │00 │FF │00 │
+ └───┴───┴───┴───┴───┴───┴───┘
+
+ NULL ┌───┬───┬───┬───┬───┬───┬───┐
+ │00 │00 │00 │00 │00 │FF │00 │
+ └───┴───┴───┴───┴───┴───┴───┘
+
+"Defenestration" ┌───┬───┬───┬───┬───┬───┐
+ │02 │'D'│'e'│'f'│'e'│FF │
+ └───┼───┼───┼───┼───┼───┤
+ │'n'│'e'│'s'│'t'│FF │
+ ├───┼───┼───┼───┼───┤
+ │'r'│'a'│'t'│'r'│FF │
+ ├───┼───┼───┼───┼───┤
+ │'a'│'t'│'i'│'o'│FF │
+ ├───┼───┼───┼───┼───┼───┐
+ │'n'│00 │00 │00 │FF │17 │
+ └───┴───┴───┴───┴───┴───┘
+```
+
+This approach is loosely inspired by [COBS
encoding](https://en.wikipedia.org/wiki/Consistent_Overhead_Byte_Stuffing), and
chosen over more traditional [byte
stuffing](https://en.wikipedia.org/wiki/High-Level_Data_Link_Control#Asynchronous_framing)
as it is more amenable to vectorization, in particular hardware with AVX-256
can copy a 32-byte block in a single instruction.
+
+### Dictionary Arrays
+Dictionary Encoded Data (called
[categorical](https://pandas.pydata.org/docs/user_guide/categorical.html) in
pandas) is increasingly important because they can store and process low
cardinality data very efficiently.
+
+A simple approach to encoding dictionary arrays would be to encode the logical
values directly using the encodings for primitive values described previously.
However, this would lose the benefits of dictionary encoding to reduce memory
and CPU consumption.
+
+To further complicate matters, the [Arrow implementation of Dictionary
encoding](https://arrow.apache.org/docs/format/Columnar.html#dictionary-encoded-layout)
is quite general, and we can make no assumptions about the contents of the
dictionaries. In particular, we cannot assume that the dictionary values are
sorted, nor that the same dictionary is used for all arrays within a column
+
+The following example shows how a string column might be encoded in two arrays
using two different dictionaries. The dictionary keys `0`, `1`, and `2` in the
first batch correspond to different values than the same keys in the second
dictionary.
+
+```
+┌ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─
+ ┌───────────┐ ┌─────┐ │
+│ │"Fabulous" │ │ 0 │
+ ├───────────┤ ├─────┤ │
+│ │ "Bar" │ │ 2 │
+ ├───────────┤ ├─────┤ │ ┌───────────┐
+│ │ "Soup" │ │ 2 │ │"Fabulous" │
+ └───────────┘ ├─────┤ │ ├───────────┤
+│ │ 0 │ │ "Soup" │
+ ├─────┤ │ ├───────────┤
+│ │ 1 │ │ "Soup" │
+ └─────┘ │ ├───────────┤
+│ │"Fabulous" │
+ Values │ ├───────────┤
+│ Dictionary (indexes in │ "Bar" │
+ dictionary) │ ├───────────┤
+│ │ "ZZ" │
+ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ┘ ├───────────┤
+┌ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ │ "Bar" │
+ │ ├───────────┤
+│ ┌───────────┐ ┌─────┐ │ "ZZ" │
+ │"Fabulous" │ │ 1 │ │ ├───────────┤
+│ ├───────────┤ ├─────┤ │"Fabulous" │
+ │ "ZZ" │ │ 2 │ │ └───────────┘
+│ ├───────────┤ ├─────┤
+ │ "Bar" │ │ 1 │ │
+│ └───────────┘ ├─────┤
+ │ 0 │ │ Logical column
+│ └─────┘ values
+ Values │
+│ Dictionary (indexes in
+ dictionary) │
+ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ┘
+```
+
+The key observation which allows us to efficiently create a row format for
this kind of data is that given a byte array, a new byte array can always be
created which comes before or after it in the sort order by adding an
additional byte.
+
+Therefore we can incrementally build an order-preserving mapping from
dictionary values to variable length byte arrays, without needing to know all
possible dictionary values beforehand, instead introducing mappings for new
dictionary values as we encounter them.
+
+```
+┌──────────┐ ┌─────┐
+│ "Bar" │ ───────────────▶│ 01 │
+└──────────┘ └─────┘
+┌──────────┐ ┌─────┬─────┐
+│"Fabulous"│ ───────────────▶│ 01 │ 02 │
Review Comment:
We don't store lengths along with the rows, in the case of dictionary keys,
they are stored null terminated. Perhaps we should update the diagram to
reflect this
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
