CDC correctness is identity plus ordering, not connector throughput

CDC (Change Data Capture) connectors are rated on throughput: how many rows per second the pipeline moves from source to target. But in a CDC system, throughput is noise unless two harder problems are solved first — identity guarantees and sequence ordering. A connector that moves 100k rows but cannot prove which row belongs to which record or in what order those changes occurred produces data that is wrong at its core.

The identity question

CDC captures change events — inserts, updates, deletes. For each event, the consumer must answer: "Does this change refer to this record?" In SQL databases, the primary key provides identity — a unique row identifier that the CDC connector copies into every change event. But in systems with composite keys, surrogate key changes, or soft deletes, identity becomes ambiguous.

The problem compounds at scale. A Debezium connector captures Postgres logical replication events; each event carries the primary key of the changed row. If the source database has no reliable primary key (or the key is changeable), the consumer receives a change — but for which record? This is not a throughput issue; it is an identity problem that makes every downstream transformation suspect.

The ordering question

CDC captures events in source database commit order. Replication buffers those events until the consumer pulls them. The consumption window creates a natural delay: changes exist in the source before they arrive at the target. If two updates to the same row occur within that window, and the consumer receives them out of order, the final state is corrupted regardless of what individual rows are correct.

Beyond per-row ordering, multi-table CDC introduces cross-record ordering. Order A updates a customer's status; Order B inserts a line item for that customer. If B arrives before A, you have a line item referencing an invalid state. The connector moved both records correctly — but the relationship between them is destroyed by out-of-order delivery.

The implication for agent-era pipelines

In traditional ETL, CDC correctness meant "the right rows arrived." In agent-era data pipelines where downstream consumers are autonomous agents making real-time decisions on the ingested data, correctness also means: "the right sequences arrived in the order the source produced them." Identity ensures you know what changed; ordering ensures you understand when it changed. Both are required for agents to act reliably on CDC state.