Reliability and recovery

Failure is an operating state. Silent loss is not.

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

Plan my migration See correctness
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.

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

Keep exploring

See the connected parts of the platform.

Correctness →

See record meaning through replay.

High Availability →

Transfer durable ownership.

Fan-out →

Use independent cursors.

Choose recovery behavior before you need it.

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

Plan my migration See correctness