Dbtrail

DBTrail keeps your MySQL tables as Parquet files you own, minutes behind the source, with every version of every row. Query them with DuckDB. Production never sees the query.

$ duckdb -init views.sql
D SELECT c.tier, count(*) AS orders, round(sum(o.total), 2) AS revenue
  FROM demo.orders o
  JOIN demo.customers c ON c.id = o.customer_id
  GROUP BY c.tier ORDER BY revenue DESC;
┌──────────┬────────┬───────────────┐
│   tier   │ orders │    revenue    │
│ varchar  │ int64  │ decimal(38,2) │
├──────────┼────────┼───────────────┤
│ platinum │ 539848 │   59426197.40 │
│ gold     │ 539752 │   59353658.54 │
│ silver   │ 539056 │   59295443.48 │
│ bronze   │ 538385 │   59242795.95 │
└──────────┴────────┴───────────────┘

Why

Reports, dashboards and ad-hoc analysis compete with your application when they run on the primary: a full scan pushes the working set out of the buffer pool, a long read holds back purge, and a GROUP BY over millions of rows spills temp tables to disk. A reporting replica is one more MySQL to pay for and patch, and it still answers at row-store speed.

DBTrail gives those queries a copy of their own, in a columnar format, outside mysqld.

How it works

1. The first copy reads your tables once (with mydumper).
2. After that, only the binlog. DBTrail connects over the replication protocol, the way a replica does. No plugin, no agent on the database host, no triggers. That is why Amazon RDS and Aurora work.
3. On your schedule, the copy is updated. DBTrail folds the recorded changes into a new Parquet snapshot, from every 5 minutes to once a day. The source is not read again, unless a table changes shape (see Limits).
4. You query it with DuckDB. A generated views.sql gives you one view per table, named like the source table (shop.orders), that follows the newest snapshot.

One Docker Compose stack, and a folder or an S3 bucket. No pipeline to build, no warehouse to run, no Kafka. See Analytics with DuckDB.

The numbers

Source: RDS MySQL 8.4, TPC-C dataset, 73 million rows, about 10 GB, under a sysbench-tpcc load of 180 transactions per second. Measured 18 to 20 September 2026 with DBTrail 0.84 refreshing every 5 minutes. Six report queries, total time:

One of the six, a two-table join grouped by district and month, took 3 min 52 s on MySQL and 0.9 s on the copy. MyDuck was faster on this set: it keeps data in DuckDB's own format, which only DuckDB reads.

Freshness, measured separately under about 340 transactions per second with a 5-minute schedule: commit to visible in DuckDB took 2.7 to 13.2 minutes, median 5.9.

Method, cost comparison and ClickBench results: Introducing the MySQL Analytical Replica.

What you get

An analytical copy

• Heavy queries off your MySQL. They run in DuckDB, in its own process, on the copy. Only the binlog leaves mysqld.
• DuckDB, ready to open. One views.sql, one view per table. Your DuckDB runs it on your machine: a laptop, a notebook, a BI box.
• Open files you own. Plain Parquet on disk or in S3, with a Hive-partitioned change history. Nothing proprietary.
• Other engines too. Export a snapshot to Iceberg for Spark, Trino or Athena with bintrail export iceberg. See Iceberg export.

Open the copy through views.sql. An engine that reads the Parquet files directly sees each table as of its last full write and can miss the change files stored beside it. The views apply them for you.

Time travel and recovery, from the same stream

DBTrail keeps every change on your MySQL server, before and after, and writes the SQL that undoes the ones you didn't want. It reads them from the same binlog as the copy, from the day you install it:

• Every version of every row. See any row or table as it was at a past moment, from the web interface or the reconstruct CLI. An optional MySQL port (beta) answers the same questions with AS OF SQL from any mysql client. See Time-Travel SQL.
• Undo precisely. Generate the SQL that reverses just the damaged rows. recover-cascade also rebuilds the child rows an ON DELETE CASCADE removed below the binary log. DBTrail writes the SQL. You review it and you run it. See Query & Recovery.
• Prove it holds. bintrail verify checks, from two snapshots and the index and without touching the source, that a recovery would reproduce it. bintrail status flags any gap the capture could not fill. See Verify.
• Web interface and MCP. Add servers, schedule snapshots and browse changes in the browser. Connect Claude Desktop or any MCP client to ask in plain English. Every tool is read-only. See the 5-minute guide.

Try time travel in 30 seconds, with nothing of yours connected:

docker run --rm -p 6033:6033 ghcr.io/dbtrail/bintrail-demo

See the demo image.

Limits

A copy you can trust is one whose edges you know.

• Not high availability. Your application never points at the copy. Nothing fails over to it.
• Not your backup. History starts the day you install it. Keep your physical backups.
• Minutes behind, not seconds, for the files. The copy refreshes on the schedule you set. Through the SQL port, a heavy read also sees what capture has indexed since, usually seconds behind.
• Needs a ROW binlog with full row images. binlog_format=ROW and binlog_row_image=FULL. bintrail doctor checks both and prints the fix (on RDS and Aurora, set them in the parameter group).
• Some changes mean a full read. Adding, dropping, renaming or retyping a column, or a TRUNCATE, DROP or RENAME of a table, pauses the scheduled update until DBTrail reads the database again, at most once a day. That read takes a short lock: a global read lock while it starts, or, on RDS and Aurora, a lock on the tables it reads. A new table joins with a read of its own. Index and other everyday changes do not trigger one.
• History needs a bucket. Without S3, row history keeps 48 hours by default (--rotate-retain changes it). The copy itself stays: the newest 3 snapshots per table on local disk.
• Cascaded deletes before MySQL 9.6, and on MariaDB. Rows deleted by a foreign key never reach the binlog. recover-cascade rebuilds most of them, and the copy keeps them until DBTrail next reads the database in full.
• Not a MySQL server. Reports run in DuckDB, not from your MySQL client. The optional MySQL port (beta, no TLS) only answers what a row or table looked like at a past moment, or what changed between two.
• Yours to operate. DBTrail and its small index MySQL run on your machines. Disk and backups are yours.

Full list: Limitations.

What it works with

DBTrail never needs the binlog files on disk, which is why managed cloud databases work.

Install

You need Docker with Compose, and DuckDB on the machine where you query.

curl -fsSL https://raw.githubusercontent.com/dbtrail/dbtrail/main/install.sh | sh

This downloads the Docker Compose stack, starts it, waits until DBTrail answers, and prints the next steps.

1. Sign in. Open http://127.0.0.1:8090, create a username and password, and press Create & sign in.
2. Connect. Click + Add server, enter the host and port of your MySQL and press Find it. The form shows the SQL that creates DBTrail's user. Run it on your MySQL, then press I ran it. DBTrail checks the rest and starts on its own.
3. First change. Change a row on your MySQL. It shows under Recent changes on the Overview within a minute, with an Undo that writes the SQL to reverse it.
4. First copy. On Snapshots, press Read database now. Normally this is the only time DBTrail reads your tables directly.
5. Set a schedule. On the Settings tab of Snapshots, under Update the copy, pick how often (every 5 minutes, for example) and press Turn on. Each run after the first folds the recorded changes instead of reading the source.
6. Query it. On Snapshots, press Download, then Download the data. Unpack the .tar.gz and run duckdb -init views.sql inside the snapshot folder. To read the copy that keeps updating, see Query in DuckDB.

Prefer the command line? See the command-line quickstart.

Who builds it

DBTrail is built by Daniel Guzman-Burgos, former MySQL Technical Lead at Percona.

Documentation

Privacy

bintrail is DBTrail's command-line tool. It runs entirely in your infrastructure. Your database data never leaves it: not your rows, schemas, table names, queries, hostnames, DSNs, or file paths.

Official release builds do report metadata-only usage statistics: which command ran, whether it succeeded, a coarse error class, and your version and platform. It is on by default and turns off in one line:

bintrail telemetry off      # or: DO_NOT_TRACK=1, BINTRAIL_TELEMETRY=off
bintrail telemetry show     # prints what would be sent; sends nothing

No identifier of any kind is stored or transmitted, so nothing ties those statistics to you or your machine. A binary you build yourself has no reporting address compiled in and cannot send anything. TELEMETRY.md documents every field, every control, and the CI tests that enforce both.

The privacy policy covers the Claude Desktop extension (.mcpb), which reports nothing at all, and how data moves when an AI client queries your deployment.

License

Apache-2.0: free for any use, including commercial and production. Contributions are welcome; see CONTRIBUTING.md (a CLA is required, prompted automatically on your first PR).

Want the index server operated for you (sized, backed up, upgraded, kept alive on-call) instead of running it yourself? That is the managed service at dbtrail.com. SUPPORT.md draws the ship-vs-operate line.