Compensating Transactions
What you'll learn: How to design idempotent undo logic that safely reverses committed operations when a saga step fails in distributed workflows.
The Challenge
When using the Saga Pattern (introduced in lesson 2721), each step commits its transaction immediately rather than waiting for a global coordinator. If a later step fails, you can't simply "rollback" — those earlier transactions are already committed! Instead, you must explicitly undo each completed step by running a compensating transaction.
Core Idea
A compensating transaction is a separate operation that semantically reverses the effect of a previously committed transaction. Think of it like a "ctrl+z" for distributed systems — but unlike a database rollback that prevents changes from becoming visible, a compensating transaction creates new changes that counteract the old ones.
Example
Imagine an e-commerce saga:
- Reserve inventory (commits)
- Charge credit card (commits)
- Ship order (fails)
To compensate, you execute:
- Un-charge the credit card (issue refund)
- Unreserve inventory (release stock)
Idempotency is Critical
Compensating transactions must be idempotent — running them multiple times produces the same result as running once. Why? Because network failures might cause retries. If "refund 100!
Solution: Check state first. Example: "If payment X exists and isn't already refunded, then refund it."
Design Patterns
- Semantic undo: Don't literally reverse bits; apply business logic (e.g., "cancel reservation" not "restore database row")
- Store compensation data: Save enough context during the forward step to enable accurate compensation later
Key Takeaway: Compensating transactions provide eventual consistency in sagas by explicitly undoing committed work through idempotent reverse operations, trading immediate atomicity for forward progress in distributed systems.