Reliability and recovery

Failure is an operating state. Silent loss is not.

SQL Flow gives every captured batch a durable address and every destination an independent cursor, with explicit recovery behavior.

Talk through my use case See correctness

Durable spool

The backlog has structure and a recovery path.

Segmented storage uses versioned frames, compression, grouped fsync, sparse indexes, and torn-tail recovery. Retention and encryption are explicit.

  • ✓Atomic append groups and source positions
  • ✓Checksummed segment recovery
  • ✓Retention, backpressure, and eviction policy
  • ✓Optional at-rest encryption

Apply resilience

Advance only after the sink contract succeeds.

Retry isolates transient failures. Binary splitting narrows a bad batch, dead letters preserve evidence, and idempotent paths absorb replay.

  • ✓Destination-specific retry policy
  • ✓Poison-record isolation
  • ✓Checkpointed pause and resume
  • ✓Explicit delivery semantics per path

The payoff

See what the capability changes.

Fsync

Durable frontier

Acknowledgement never outruns persisted changes.

N cursors

Independent recovery

Each sink restarts from its own progress.

Explicit

Failure policy

Retry, dead-letter, backpressure, or reseed.

Take the next step

Choose recovery behavior before you need it.

Bring outage windows, retention, target idempotency, and failure policy. We will map recovery.

Talk through my use case