Boosting MySQL with Vector Search Scale22X 2025.pdf

Boosting MySQL with Vector Search
Scale22X 2025
Introducing the MyVector Plugin
Alkin Tezuysal

Let’s get connected!
Alkin Tezuysal - Director of Services @AltinityDB
● Linkedin : https://www.linkedin.com/in/askdba/
Open Source Database Evangelist
● Previously ChistaDATA, PlanetScale, Percona and Pythian as Senior Technical Manager, SRE, DBA
● Earlier in life Enterprise DBA , Informix, Oracle, DB2 , SQL Server
● Recent Recognitions:
○ Most Influential in Database Community 2022 - The Redgate 100
○ MySQL Cookbook, 4th Edition 2022 - O'Reilly Media, Inc.
○ MySQL Rockstar 2023 - Oracle (MySQL Community)
○ Database Design and Modeling with PostgreSQL and MySQL 2024 - <Packt>
@ask_dba
@2025 Altinity, Inc.
Born to Sail, Forced to Work!

Sailing Trivia
What is the name of the horizontal spar that
extends from the mast to control the angle
of the sail?

Vectors and Dimensions
● Mathematical objects representing
magnitude and direction (velocity,
force)
● Vectors can be represented as
arrays or ordered lists of numbers
○ 2D (x, y), 3D (x, y, z), n-dimensional
(AI)
● Basically data points
● When dimensions increase data
points increase and they turn into
embeddings

Understanding Vector
Embeddings
● Vector Embeddings: Mathematical
representations of data in a
high-dimensional space.
● Capture semantic meaning and
relationships between words, phrases, or
other data.
● Enable more accurate and relevant search
results compared to traditional keyword
search.
● Power AI applications like semantic search,
recommendation systems, and chatbots.
● Data analysis and application functionality

Vector Embeddings
Transforming data into vectors in
a way that captures the semantic
meaning or relationships between
the data points. (The image on the
left shows a OpenAI vector
embedding for a Wikipedia article
using text-embedding-ada-002
Vector dimension is 1536)
OpenAI has another model text-embedding-3-large model provides
3,072-dimensional vector embeddings. P.S: Deep Seek can
generate embedding dimensions upto 8K!

Benefits of Vector
Search
● Efficient and relevant search results.
● Improved relevance for text and image
search.
● Semantic understanding.
● Supports AI-based applications.
● Leverage and unlock value from live
and historical business data in the
MySQL database : e.g conversation
logs, clinical records, support tickets,
persona descriptions, medical images,
product catalog, legal filings etc.

Vector Support Helps
● Traditional search struggles with
high-dimensional data and semantic
understanding.
● Inefficient for AI and ML applications
that rely on vector embeddings.
● Vector support enhances performance
and relevance for similarity searches.
● Enables new AI-driven applications
directly within MySQL.
● Unlocks advanced data analysis
techniques using vector embeddings.

Why Vector Search in MySQL?
● MySQL World’s number one open-source
relational database
● Vector search allows for more efficient and
relevant search results.
● Improved relevance for text and image search
with semantic understanding.
● MyVector plugin extends MySQL with vector
search capabilities.
● Supports a wide range of AI-based applications.
P.S: MySQL does not come with extensibility
interfaces to add new data types, and no support
to add new index types and to add new access
methods

Vector Search Use Cases
Use Cases:
Natural Language Processing (NLP): Words or
sentences are converted into vectors where similar
words or phrases are closer in vector space.
Recommender Systems: Items or users can be
embedded to find similarities or predict preferences.
Image Recognition: Images can be encoded into
vectors for tasks like similarity search or classification.

Vector Databases
Vector Database
Providers
● Pinecone
● Qdrant
● Milvus
● Weavite
● Chroma
Vector Store
Providers
● ClickHouse
● Elasticsearch
● Singlestore
● Typesense
● Others
○ MySQL
○ MariaDB
○ PostgreSQL
○ TiDB.

Vector Databases vs Traditional Databases

Vector Data Processing

MySQL Vector Plugin Architecture
The MyVector Plugin, discussed earlier,
leverages this architecture. Here's how it
fits:
● UDFs in the plugin: Adds vector-related
SQL functions.
● Data storage : Use varbinary in MySQL
8.x and VECTOR in MySQL 9.x.
● Indexing: Integrates vector similarity
search via HNSW indexing.
● Similarity Search - Intuitive syntax for
search using SQL
● Administration - Simple MySQL stored
procedures

MySQL Vector Plugin Build
/// get the MyVector sources
$ cd mysql-server/src/plugin
$ git clone https://github.com/p3io/myvector-dev/ ./myvector
/// Generate makefile for the new plugin
$ cd mysql-server/bld
$ cmake .. <other options used for this build >
/// Build the plugin
$ cd mysql-server/bld/plugin/myvector
$ make

MySQL Vector Plugin Install
/// Copy the MyVector plugin shared library to the MySQL installation
plugins
$ cp mysql-server/bld/plugin_output_directory/myvector.so
/usr/local/mysql/lib/plugin/
/// Register the MyVector plugin and create MyVector stored procedures.
$ cd mysql-server/plugin/myvector
/// Connect to 'mysql' database as 'root'
$ mysql -u root -p mysql
mysql> source myvectorplugin.sql

MySQL Vector Plugin Architecture
● MySQL V8.X - InnoDB + Plugin + Varbinary Data Type
● MySQL V9.X - InnoDB + Plugin + Vector Data Type
mysql> select udf_name from performance_schema.user_defined_functions
where udf_name like "myvector%" ;
+-----------------------------+
| udf_name |
+-----------------------------+ |
| myvector_is_valid |
| myvector_distance |
| myvector_display |
| myvector_ann_set |
| myvector_row_distance |
| myvector_construct |
+-----------------------------+

MyVector Features
High speed, parallel build of HNSW index
Specify recall v/s latency parameters : M, ef, ef_search
HNSW index is incrementally persisted after initial build
Online update of index via binlog read & parse of DMLs
HNSW Index is crash-safe and is recovered on MySQL
instance crash
Write amplification to redo log/undo log avoided

MySQL Vector Plugin Examples MySQL 8.X
Ref: https://nlp.stanford.edu/projects/glove/
mysql > create table words50d( wordid
int primary key,
word varchar(200),
wordvec MYVECTOR(type=HNSW,
dim=50,size=400000,
M=64,ef=100)
);
– Load the Data (insert …)
mysql> call
myvector_index_build
('test.words50d.wordvec','wordid');

MySQL Vector Plugin Examples MySQL 9.1
mysql > create table words50d(
wordid int primary key,
word varchar(200),
wordvec MYVECTOR(type=HNSW,
dim=50,size=400000,
M=64,ef=100)
);
– Load the Data (insert …)
mysql> call
myvector_index_build
('test.words50d.wordvec','wordid');

MySQL Vector Plugin Examples
myvector_construct(vector_string VARCHAR):
Purpose: Converts a human-readable vector string into a serialized binary format suitable for storage
in a VARBINARY / VECTOR column.
Usage Example:
INSERT INTO vectors_table (vector_column)
VALUES (myvector_construct('[0.1, 0.2, 0.3, ...]'));

myvector_display(vector_col_expr VARBINARY/VECTOR):
Purpose: Transforms a binary-stored vector back into a human-readable string
representation.
Usage Example:
SELECT myvector_display(vector_column) AS readable_vector
FROM vectors_table;

myvector_distance(vec1 VARBINARY/VECTOR, vec2 VARBINARY/VECTOR, disttype VARCHAR):
Purpose: Calculates the distance between two vectors using the specified distance metric ('L2',
'EUCLIDEAN', or 'IP' for inner product).
Usage Example:
mysql> select myvector_distance((select wordvec from words50d where word =
'school'), (select wordvec from words50d where word='institute'))G
*************************** 1. row ***************************
myvector_distance((select wordvec from words50d where word = 'school'), (select
wordvec from words50d where word='institute')): 25.51254653930664
1 row in set (0.36 sec)

Similarity Search (ANN)
SELECT <column-list> FROM <table> WHERE
MYVECTOR_IS_ANN('vector_column','key column',
<search vector>, options)
// search vector(s) should first be inserted to <query_table>
SELECT <column-list> FROM
MYVECTOR_SEARCH[<base_table>, <key column>,
<vector_column>, <query_table>]

myvector_ann_set(veccol VARCHAR, options VARCHAR, searchvec
VECTOR/VARBINARY):
Purpose: Returns a comma-separated list of IDs corresponding to the nearest neighbors
of a given search vector.
Usage Example:
SELECT myvector_ann_set('vector_column', 'options',
myvector_construct('[0.1, 0.2, 0.3, ...]')) AS neighbors FROM
vectors_table;

Vector Index Administration - Stored Procedures
// build the index
CALL myvector_index_build('vector column', 'key column')
// check current status of the index - number of rows, parameters etc
CALL myvector_index_status('vector column')
// open/load the index into memory after a restart (if not marked
online)
CALL myvector_index_load('vector column')
// drop the the index
CALL myvector_index_drop('vector column')

Initial Benchmarking Results
Dataset : gist-960-euclidean

Initial Benchmarking Results
Dataset : dbpedia-openai-1000k-angular

Demo
https://github.com/p3io/myvector-dev/tree/main/demo

References
Introducing MyVector Plugin — Vector Storage & Similarity Search in MySQL - HNSW, ANN | Medium
Introducing MyVector Plugin — Vector Storage & Similarity Search in MySQL — Part 2 | by Shankar Iyer | Medium
https://medium.com/@shiyer22/myvector-plugin-part-3-nearest-neighbour-search-with-other-predicates-and-the-sql-o
ptimizer-7541bfaa3df9
MySQL Vector Datatype: create your operations (part 1)
MySQL Vector Datatype: create your operations (part 2) - lefred blog
O'Reilly/Safari Books
https://asciinema.org/a/KZeSifbD0QpHYmYUep5SiaCMA
https://asciinema.org/a/YTIg12ClZAh38qUcgsBUAdSA7
https://asciinema.org/a/4uro9wYgFijuwP3EdJvp0tRMj
https://asciinema.org/a/O7rNs2OzLXyUja0bwWcldXsX9
https://asciinema.org/a/RJs3CFbuShQGeXkCSxDDG0n3i
https://asciinema.org/a/6F0hDcnFT4XMqi5ALSeTNkCbO
https://fosdem.org/2025/schedule/event/fosdem-2025-4230-boosting-mysql-with-vector-search-introducing-the-myvect
or-plugin/

Boosting MySQL with Vector Search Scale22X 2025.pdf

Recommended

More Related Content

More from Alkin Tezuysal (20)

Recently uploaded (20)

Boosting MySQL with Vector Search Scale22X 2025.pdf