Code, Vectors & Vision: A MariaDB Meetup with the Founder

MariaDB: Codes Vectors & Visions 2025-05-06, Bremen (Germany)

Get to know MariaDB’s Blazing-Fast Vector search

Slides from Sergei Golubchik (MariaDB Foundation) hopefully here…

MariaDB Vector Tuning

Tuning is hard

• Many (> 10) algorithm parametrised.
• Impossible to describe what they do.
• Complex interaction between parameters.
• MariaDB solution: Hide that and auto-tune!

SQL> SHOW GLOBAL VARIABLES LIKE 'mhnsw%';
+------------------------+-----------+
| Variable_name          | Value     |
+------------------------+-----------+
| mhnsw_default_distance | euclidean |
| mhnsw_default_m        | 6         |
| mhnsw_ef_search        | 20        |
| mhnsw_max_cache_size   | 16777216  |
+------------------------+-----------+

Vector System Variables

mhnsw_max_cache_size

• HNSW is in-memory algorithm.
• Otherwise slow.
• At least as big as half of the size of the data.

• Set as large as possible.
• Grows as needed up to the specified limit.

DISTANCE mhnsw_default_distance

• euclidean or cosine
• Depends on the problem area
• Practically:

→ cosine (Text) → Images euclidean

Depends on how the vectors were generated.

mhnsw_ef_search and mhnsw_default_m

• slow and precise ⇔ fast an imprecise
• fast write / slow search ⇔ slow write / fast search

mhnsw_ef_search affexts only the search.

mhnsw_default_m affects both write and search.

build time vs. recall

M=3..5 ok

M=6 starts exponential growth

M=12 too high

qps vs. recall

Better to increase mhnsw_default_m than mhnsw_ef_search.

Contributing to MariaDB

Oleksandr “Sanja” Byelkin

Bug reporting

→ https://jira.mariadb.org

A bug should be:

• Repeatable
• Repeatable in simple steps
• Repeatable with a small test case
• Repeatable and with proposed workaround
• Repeatable and with propsed fix

Non-programming contribution

• Zulip
• Knowledge base: https://mariadb.com/kb/en/
• Error messages translation (a lot of work) https://github.com/MariaDB/server/tree/main/sql/share

Contributing code

• Google summer of code
• Breakpoint: my_message_sql(). See also Marcos comment