sagas-garcia-molina-salem-1987

Distributed transaction recovery for long-running workflows uses a compensation-based pattern with checkpointing save-points and externalized in-transit state, enabling both backward recovery through compensating actions and forward recovery through retries to maintain semantic consistency.

Designing compensation for multi-step agent orchestration involves decomposing tasks into discrete saga steps and specifying a compensating action for each step, ensuring partial executions can be semantically undone without leaving external systems in inconsistent states.

Long-running workflow execution leaves inconsistent partial states because resource blocking, irrecoverable side effects, and deadlock exposure increase exponentially as task scope grows, making partial execution visible to other processes without compensation handling or save-point placement.

Externalize in-transit state for distributed workflow observability by applying the funds-in-transit principle, which surfaces intermediate state externally so workflows become observable, recoverable, and auditable by other processes during partial execution.

Use backward recovery with compensating actions when partial side effects must be semantically undone, and use forward recovery with checkpoint-and-retry when transient failures occur; mixed recovery combines both approaches based on save-point placement guidelines.

Yes, the saga daemon pattern wraps underlying systems with instrumentation rules to orchestrate recovery without native support, enabling scheduled batch jobs and large-data transformations to remain resilient to crashes through externalized state and compensation.