> ## Documentation Index
> Fetch the complete documentation index at: https://filament.getgalaxy.io/llms.txt
> Use this file to discover all available pages before exploring further.

# MySQL

> Write to MySQL.

`connectors/mysql/sink` loads each resource into its own typed table with
native MySQL columns.

## Typed loading

The engine calls `EnsureSchema` once per resource before extraction, which
creates the destination table (adding any new columns to an existing one)
and caches a prebuilt `INSERT` statement. `Write` then casts each batch's
JSON row payloads into those typed columns server-side via `JSON_TABLE`
(MySQL 8.0+) — the same role Postgres's sink gives `jsonb_to_recordset`. A
batch never gets parsed client-side; it's framed as one JSON array and
handed to MySQL as a single bound parameter.

Binary columns round-trip through `TO_BASE64`/`FROM_BASE64` (MySQL JSON has
no binary representation, matching the source's encoding convention); JSON
columns pass through unchanged so nested structure survives the round trip.

## Write policies

A resumable ingestion type (`SnapshotUpsert`, `Upsert`, `CDC`) upserts by
primary key via `ON DUPLICATE KEY UPDATE`, so a re-delivered row from an
at-least-once resume overwrites rather than errors on the unique
constraint. A non-resumable load skips the upsert clause and truncates the
table first for full-snapshot replace semantics.

CDC batches apply through a different path: consecutive same-op records
(insert/update vs. delete) are split into maximal runs in arrival order —
so a delete never jumps ahead of the insert it's paired with — and each run
lands as one statement: inserts/updates through the same upsert `INSERT …
JSON_TABLE`, deletes through a `JSON_TABLE`-join `DELETE` keyed on the
batch's primary-key columns.

## Modes

Supports `SnapshotReplace`, `SnapshotUpsert`, `Append`, `Upsert`, and `CDC`
write policies (see [Replication
modes](/pages/guides/concepts/replication-modes)).