# Senior CDC / streaming-ETL playbook

Compressed reference for the senior-only CDC extension. Pair with [cheat-sheet.md](./cheat-sheet.md), [GLOSSARY.md](./GLOSSARY.md), and [repo-cdc-map.md](./repo-cdc-map.md).

## The senior shift
Intermediate understanding is: "I know the WAL → Debezium → Kafka → sink/ksqlDB → destination flow, the three stacks, generated connectors, and hephaestus reconcile."

Senior understanding is: "I can reason about replay safety, schema drift, slot-safe backfills, no-DLQ trade-offs, and whether a CDC design is making truthful claims about continuity, delivery, and operations."

## Six senior lenses
1. **At-least-once is not enough** — sink idempotency is part of the guarantee.
2. **Schema evolution is a pipeline problem** — not only a table-migration problem.
3. **Backfills are slot-safety stories** — the source path is continuity-sensitive.
4. **No-DLQ is a real trade-off** — simplicity and recovery strength move together.
5. **Generated fleets reduce drift, not judgment** — the human still owns semantics.
6. **Review is about guarantee truthfulness** — ask what the design really guarantees and what it hides.

## Senior review checklist
- Can you explain why this flow is at-least-once but still safe under replay?
- Can you distinguish source-recreation risk from sink churn clearly?
- Can you describe schema evolution across source, topic, sink, and warehouse together?
- Can you explain what no-DLQ buys and what it gives up?
- Can you justify generation + reconcile as a control-plane choice for this connector fleet?
- Can you say when CDC is the right fit and when outbox/direct domain events may be a better answer?

## Interview answer shapes
### Why isn’t at-least-once the full answer?
Because the transport can repeat work; the sink’s idempotent upsert is what makes repeating that work safe.

### Why is source recreation riskier than sink recreation?
Because the source owns continuity with the WAL through the replication slot, while a sink is only a downstream applier of derived data.

### What is the real value of incremental snapshots?
They let you backfill existing rows for new tables while preserving the running source connector and its slot continuity.

### What is the real point of hephaestus?
It makes git the desired state for a large connector fleet and converges live Connect clusters safely through reconcile and guarded deletion.

### What is the real job of a senior CDC reviewer?
To ask what exact guarantees the design provides about continuity, replay, schema safety, and failure handling, and whether the trade-offs match the implementation.