How I handle multi-step rollbacks in Go without external infrastructure

Semyon — Sun, 22 Feb 2026 06:31:08 +0000

The problem

Every non-trivial backend service has operations that span multiple steps. A classic example - placing an order:

Charge the customer's card
Reserve items in the warehouse
Create a shipment

Simple enough. But what happens when step 3 fails? You need to release the reservation and refund the charge. In the right order. With the right data.

Most teams handle this by hand. It looks something like this:

func PlaceOrder(ctx context.Context, req *Request) error {
    chargeID, err := payments.Charge(ctx, req.CardToken, req.Amount)
    if err != nil {
        return err
    }

    reservationID, err := warehouse.Reserve(ctx, req.ItemID)
    if err != nil {
        _ = payments.Refund(ctx, chargeID)
        return err
    }

    if err := shipping.Create(ctx, reservationID); err != nil {
        _ = warehouse.Release(ctx, reservationID)
        _ = payments.Refund(ctx, chargeID)
        return err
    }

    return nil
}

This works. Until it doesn't.

Add a fourth step and you touch every error branch. Forget one _ and you have a silent bug. The compensation logic is scattered across the function instead of being colocated with the step it compensates.

Why not Temporal?

Temporal is a great tool - for the right problem. But it requires running a dedicated server cluster and introduces significant operational complexity for what is essentially a local coordination problem.

If you're already running Temporal, great. But if you just want clean rollback logic in a Go service without spinning up new infrastructure - it's overkill.

A cleaner approach

I wanted something that works like this:

runner := kata.New(
    kata.Step("charge-card", chargeCard).
        Compensate(refundCard).
        Retry(3, kata.Exponential(100*time.Millisecond)),
    kata.Step("reserve-stock", reserveStock).
        Compensate(releaseStock),
    kata.Step("create-shipment", createShipment),
)

if err := runner.Run(ctx, &OrderState{
    CardToken: req.CardToken,
    Amount:    req.Amount,
    ItemID:    req.ItemID,
}); err != nil {
    // refundCard and releaseStock already ran automatically
}

If create-shipment fails, the library automatically calls releaseStock then refundCard - in reverse order, with the full shared state available to each compensation.

No infrastructure. No DSL. Just Go.

Key design decisions

Shared typed state instead of a chain

Each step reads from and writes to a shared struct. This means compensations always have access to IDs and data created by earlier steps - which is exactly what you need for a real refund or release.

type OrderState struct {
    CardToken     string
    Amount        int64
    ChargeID      string // filled by charge-card step
    ReservationID string // filled by reserve-stock step
}

func chargeCard(ctx context.Context, s *OrderState) error {
    id, err := payments.Charge(ctx, s.CardToken, s.Amount)
    s.ChargeID = id
    return err
}

func refundCard(ctx context.Context, s *OrderState) error {
    return payments.Refund(ctx, s.ChargeID)
}

Generics for type safety

The runner is generic over your state type - no interface{}, no casting:

var orderRunner = kata.New(
    kata.Step("charge-card", chargeCard).Compensate(refundCard),
    // ...
)

Two distinct error types

Not all failures are equal. If a step fails and all compensations run successfully - that's a clean rollback. If a compensation also fails - that's a potential data inconsistency requiring manual intervention.

var stepErr *kata.StepError
var compErr *kata.CompensationError

switch {
case errors.As(err, &stepErr):
    log.Printf("step %q failed: %v", stepErr.StepName, stepErr.Cause)

case errors.As(err, &compErr):
    pagerduty.Fire(compErr)
}

Parallel steps

Sometimes you want to run steps concurrently - like sending email, SMS, and push notifications at the same time:

kata.Parallel("notify",
    kata.Step("email", sendEmail),
    kata.Step("sms", sendSMS).Compensate(cancelSMS),
    kata.Step("push", sendPush),
)

If any step in the group fails, the others are cancelled and successful ones are compensated.

What it doesn't do

This library is intentionally scoped. It does not:

Persist state to a database (no crash recovery)
Coordinate across services over a network
Replace Temporal for long-running workflows

If you need those things - use Temporal. kata is for in-process coordination where you want clean rollback logic without the operational overhead.

Try it

go get github.com/kerlenton/kata

GitHub: https://github.com/kerlenton/kata

Zero dependencies, requires Go 1.22+. Still early - feedback on API ergonomics especially welcome.