Plenty of jobs come in steps that have to finish together or not at all. An order reserves stock, charges a card and books a courier. A sign-up creates an account, sets up billing and sends a welcome email. A device update downloads, verifies and installs. The risk is always the same: the process dies halfway, and you’re left with a charged card and no courier.
There are two usual answers. A status column with hand-written retry code, which works until two edge cases meet. Or a workflow platform, which works well but brings a server, a database and a second system to run, upgrade and monitor, sometimes bigger than the program it serves. FlowChassis sits between them: a small workflow execution kernel for Go that runs your steps inside your own program and keeps its state in one journal file.
The Pitch in Four Numbers
- 240 kills across two recorded crash demos, 1,200 orders: 899 completed, 301 compensated, none lost.
- Every byte of the journal’s last record cut off in turn, and every journal opened and resumed.
- 26,000 steps a second with 64 workers, with every checkpoint flushed to disk.
- 0 modules needed besides the Go standard library. No server, no database.
How It Works
A flow is a list of named steps. A run moves to the next step, or to whichever one a step picks. Before a step runs and after it succeeds, FlowChassis writes a record to the journal and flushes it to disk. Records written at the same moment share one flush, which is how it stays fast.
When a step fails, it retries after growing waits, up to a limit you set. The waiting is journaled too, so a restart doesn’t cut it short. If a run can’t finish, the undo actions of the steps that did complete run in reverse order, newest first, each seeing the state its own step left behind. That’s the saga pattern, done for you.
After a crash, the engine reads the journal and carries on. A step that was interrupted runs again. A step that finished doesn’t. In the site’s example, a process killed during “charge” restarts at charge and skips the reserve step it already did. The flowchassis command lists runs, shows their history, checks a journal record by record and compacts it.
The README is the manual, and the repository includes a contract document that spells out the engine’s guarantees and a specification of the journal format. Testing is about failure: workers killed hundreds of times mid-flow, journals cut at every byte, disks filled during runs, on Linux, macOS and Windows with every change.
Who It’s For
- Go developers with order flows, sign-ups, provisioning or payments that must not end half done.
- Teams that don’t want to run a workflow platform just to get reliable multi-step jobs.
- Embedded and edge software where a second server isn’t an option.
- Product builders who want to embed a workflow engine in their own software.
Limits
One process per journal on one machine. Steps run at least once, so steps that call outside services should use the idempotency key the engine provides. There’s no graphical designer, dashboard, server or cluster, on purpose. Those can be built on top of a kernel like this one.
The Four-Question Screening
Every project featured here gets the same four questions, the way you’d screen a startup idea.
1. Where is the money in it?
FlowChassis is open source under Apache 2.0, with no paid tier. Workflow engines are a real and growing market, mostly sold as platforms. FlowChassis’s money is in embedding: software vendors who want durable workflows inside their own product, without shipping a second system to customers, pay for support and a commercial relationship. The site’s only commercial call is exactly that, a contact for embedding. A paid layer on top, dashboards, multi-machine coordination or a hosted control plane, is a later step.
2. What real problem does it solve?
Half-finished jobs. They lead to charged cards without orders, accounts without billing, devices stuck mid-update, and they’re painful to clean up by hand. Getting the retry and undo logic right is harder than it looks, which is why so many teams get it wrong.
3. The technology in plain words
Think of a cook following a recipe who ticks off each step on paper as they go. If the power cuts out, they look at the paper when it comes back and carry on from the last tick, without adding the salt twice. And if they find halfway through that they can’t finish the dish, they put everything back in the order they took it out.
4. Moat and differentiation
Big workflow platforms are powerful and heavy. Hand-rolled retry code is light and fragile. FlowChassis offers durability as a library with one file, backed by a written contract and crash tests that show the numbers. That’s rare. The moat that can grow is the reputation of the guarantees, which is what an embedding vendor buys, and the case studies that show how to build real flows on it.
Why It Matters
Every serious application has a few multi-step jobs that must not end half done. Making them reliable shouldn’t need a second system. A small kernel that just works, and proves it, lowers the bar for doing it right.
Follow the Project
Read what a workflow execution kernel is and why sagas need one, start with the quick start, see how it was tested with 240 worker kills, or get the code from the FlowChassis repository on GitHub. FlowChassis is listed with the other open-source projects on SoftDevStudio.com, and more featured projects are in the Projects section.