← AWS Solutions Architect Professional: complex decisions
13 / 25 · 75 MIN

DMS replication and cutover evidence

Distinguish capture, apply, and validation when deciding whether data is ready for transition.

Connect initial state to the change stream

Full load and CDC solve different parts of migration. A CDC-only path assumes the target already has suitable initial state; it does not automatically fill omitted historical data. Connect the initial snapshot to the change-stream starting point and demonstrate that no interval is uncovered. In Oracle, a native point can be an SCN; mechanisms and prerequisites vary by engine. Also retain the recovery checkpoint of a stopped task before deleting it, because that context can be lost. In the fictional exercise, the snapshot ends at 01:00 but CDC starts at 01:12. The PM should not accept twelve untreated minutes merely because the task is Running. Request continuity evidence or a validated reload/reconciliation procedure for the interval. A readable timestamp alone does not establish that required logs remain available.

Locate delay before resizing

Compare CDCLatencySource with CDCLatencyTarget over the same interval. If both rise similarly, start with source capture. If source latency stays near zero while target latency grows, investigate apply, indexes, resources, and the path to the target. This guides investigation; it does not identify one unique cause by itself. A long uncommitted transaction can increase the source metric. CDCIncomingChanges represents captured changes awaiting apply, not the rate of new transactions. In the rehearsal worksheet, record metrics, units, task state, logs, and business load. Avoid concluding that more CPU fixes it before locating the constraint. During closing, additional validation load or parallelism can compete with the original application, so every adjustment needs observation and stopping criteria.

Validate the data that will actually be used

Do not treat every table as equivalent. DMS validation requires a compatible primary key or unique index; in Oracle, a NOVALIDATE key is not accepted as that guarantee. Continuously modified rows can remain unvalidated. Target writes outside DMS during checking can make discrepancy reporting inaccurate. Identify these conditions in the rehearsal plan and distinguish a tool limitation from an actual data difference. For LOBs, limited mode truncates values above the configured maximum and issues a warning. A sample containing only small documents can hide the problem. Measure the distribution, include representative extremes, and choose a supported configuration preserving required data. Inline mode handles small values inline and larger values through full LOB where the endpoint supports that mode.

Prepare the source and decide transition

For Oracle CDC, check database-level and table-level supplemental logging, log access, and migration-user permissions for the selected method. A successful connection does not establish correct capture of updates and deletes. Rehearsal should include representative inserts, updates, deletes, and LOBs, with owner-defined expected results. Before authorizing target writes, assemble evidence of complete scope, treated pending changes, explained differences, and APS access. Also define who can pause producers, how to identify the last accepted business set, and how to retain writes made after transition. CDC latency varies and should not be presented as real-time replication with an invented SLA. Go/no-go combines technical evidence and business acceptance; a green task is only part of the dossier.

Original DMS triage worksheet; not an AWS command
CDCLatencySource = 2 seconds
CDCLatencyTarget = 240 seconds
CDCIncomingChanges = 180000 pending changes
First investigation: target apply path
Not proven: exact cause, business reconciliation, cutover readiness
IN PRACTICE

Source latency of two seconds, target latency of 240, and 180,000 pending changes: investigate target apply before concluding capture is slow.

Common pitfalls

Confusing Running with completeness; deleting a task before retaining its checkpoint; omitting large LOBs; using average latency as proof of no gaps.

Related topics: File and interface migration

Take this idea with you

Capture, apply, and validation each need evidence and a defined cutover boundary.

Create account

Reference: DMS ongoing replication · SAP-C02

AWS is a trademark of Amazon.com, Inc. or its affiliates. bigsavant.com is an independent preparation platform and is not affiliated with, associated with, sponsored, authorised or endorsed by AWS. Content and questions are original, are not official exam questions, and completing our tests does not award or guarantee any certification. Names are used only to identify the subject. All other trademarks belong to their respective owners.