Atrium Tests Snowflake MySQL CDC: 20-30s Latency, Cascade-Delete Gap, Preview Status

Snowflake’s Connector for MySQL kept a test Snowflake database current with a live MySQL server in about 20 to 30 seconds per change, per a hands-on evaluation published by Atrium, which also flagged a cascading-delete gap and unclear general-availability prospects for the Preview feature.

That latency figure matters because it came from an actual run, not a marketing deck. Atrium built a MySQL 8 database in Docker, loaded Sakila — MySQL’s reference dataset for a fictional DVD rental shop: 599 customers, 1,000 films, 16,044 rental records — and pointed the connector at it.

The initial load finished in a couple of minutes with row counts matching on both sides. Then came the real test: change data capture, or CDC, where the destination tracks individual changes rather than re-copying whole tables.

The mechanics are worth understanding before anyone signs a migration contract. MySQL already writes every insert, update, and delete, in order, to its binary log — the binlog. A CDC tool replays that log downstream instead of diffing tables.

Snowflake’s implementation is a two-part design: a native application, installed from the Snowflake Marketplace, manages configuration, replication state, and the destination tables; an Agent runs near the database — on-premises, a cloud VM, or in this test a second Docker container — reads the binlog, and pushes changes to Snowflake.

Snowflake’s Agent documentation says one Agent can serve as many MySQL data sources as it can reach directly, which means a central Agent can cover several databases without new infrastructure on every server.

The handshake is deliberately simple. Snowflake generates a configuration file with a temporary credential; on first startup the Agent exchanges it for a permanent key pair. In Atrium’s run, the Agent checked in on a regular cycle and picked up newly registered tables without a restart.

Three changes, one verdict

Status indicators prove nothing on their own, so the test made three changes in MySQL and watched for them in Snowflake. An update to a customer’s name appeared with no manual trigger. A new customer insert was read from the binlog and replicated on its own.

The third — deleting a customer with existing payment records — was rejected by MySQL at the source, the same InnoDB foreign-key constraint any application would hit, so no change event was generated and the connector correctly replicated nothing.

A batch connector would have picked up the first two only on its next scheduled run, potentially hours later. Twenty to thirty seconds is competitive with dedicated CDC tooling in the same class, though Atrium cautions that production latency depends on change volume, network, and configuration — measure it in your own environment.

There’s a wrinkle that belongs in every architecture review. MySQL’s ON DELETE CASCADE removes child rows automatically when a parent is deleted, and those cascaded deletions never hit the binlog.

Snowflake documents that for the Openflow Connector for MySQL, which reads the same log, so treat it as a constraint for either connector: only the parent row’s deletion reaches Snowflake, and dependent rows stay behind. Anything built on cascading deletes needs a cleanup strategy before production.

The journal hides a freebie

Here’s the part I find genuinely clever. Underneath the tables you query, the connector maintains a JOURNALS schema logging every insert, update, and delete since replication started, in order. Your day-to-day tables carry only the latest version of each row — a Type 1 slowly changing dimension, fine for operational reporting.

When someone asks what a row looked like on a given date, the journal means Type 2 history becomes a modeling job on data you already hold, not a new pipeline. The connector doesn’t build the Type 2 table for you, and nothing before replication began is captured — start CDC early, because most teams discover they need history after it’s gone.

ENUM and SET, two MySQL types with no Snowflake native equivalent, both mapped to VARCHAR without fuss. Four tables — customers, films, rentals, payments — moved from schema introspection through snapshot load into continuous CDC, tracked via a REPLICATION_STATE query showing DONE and IN_PROGRESS stages.

Openflow connector replaces preview with four key checks

Now the uncomfortable part. General availability for the connector Atrium tested is currently not on Snowflake’s roadmap. Snowflake directs new MySQL replication work to the Openflow Connector for MySQL, which is generally available, replicates MySQL 8.0, 8.4, and MariaDB 11.4 or later, and moves tables through the same three stages. Openflow is configured differently and has its own limits — rerun the test there before anything ships to production.

Per Atrium’s evaluation, four checks matter most for an Openflow implementation:

  • Timing — which tables need near-real-time changes versus a scheduled refresh.
  • Capacity — sizing the Openflow runtime for sustained change volume.
  • Row identity — every table needs a primary key or a NOT NULL unique index; tables without one need a logical key configured.
  • Deletes — Openflow does not replicate rows removed by a cascade and ignores TRUNCATE, so those rows would linger in Snowflake.

The broader read: the native-app-plus-Agent pattern gives Snowflake an answer to the market — which has built its business on exactly this binlog-replay plumbing — while Openflow handles the general-purpose integration load.

A connector that replicates a clean Sakila test in under half a minute is a proof of concept; the roadmap referencing the Preview connector is a business decision. A one-database win is only the first data point for a migration spanning many source systems, each with its own change frequency and lag tolerance.

The gap between a working demo and an architecture a business can depend on is precisely those four checks. Everything else was, by the test’s own numbers, already fast.

Photo of author

Technology Desk

Technology Desk at Archyde.

Dow Falls 500+ Points on Rate Fears as Anthropic IPO Prices Above Range and Nvidia Sinks