A safe Salesforce data migration runs nine checks: agree the scope and history cut-off, profile the source for quality and duplicates, agree a written field map, prepare the target org, load parents before children using external IDs, dry-run the full volume in a sandbox, reconcile counts and totals against the source, run a rehearsed cutover, and keep a written rollback plan.
Key takeaways
- Profile before you map. You cannot map data you have not measured.
- Every migration gets a full-volume sandbox dry run and a row-level reconciliation.
- Carry legacy IDs into external ID fields so relationships and history keep resolving.
- Load in dependency order: users, accounts, contacts, then the records that point at them.
- A cutover without a rollback plan is a hope, not a plan.
Most failed Salesforce go-lives are not caused by configuration. They are caused by data that arrived late, arrived wrong, or arrived without the links that made it useful. This is the checklist we work through on every implementation, whether the source is a spreadsheet, HubSpot, Zoho, Dynamics or an older Salesforce org.
The checklist at a glance
| Phase | What to check | Done when |
|---|---|---|
| Scope | Objects in scope, history cut-off date, systems that stay live | Signed scope list with an archive plan for the rest |
| Profile | Record counts, fill rates, duplicates, orphaned children, picklist values | Profiling report shared with the data owners |
| Map | Every source field, destination, transformation and default | Field map approved by the business owner of each object |
| Prepare the org | External ID fields, record types, picklist values, users, audit-field permission | Target org passes a test insert of every object |
| Dry run | Full volume into a full or partial sandbox, in production load order | Zero unexplained errors and a timed run |
| Reconcile | Row counts, sum totals, relationship checks, spot checks | Signed reconciliation report |
| Cutover | Source freeze, delta load, automation off and on, smoke test | Go/no-go signed by the named decision-maker |
| Rollback | What reverses, what stays, who decides, by what time | Written plan rehearsed in the sandbox |
1. Agree the scope and the history cut-off
Before profiling, decide which objects move and how far back. Five years of closed cases and every email ever logged rarely earn their storage cost or their load time. We usually migrate open records plus two to three years of history, and archive the rest somewhere searchable. Write the cut-off down, because it will be questioned during user acceptance testing.
Also list the systems that stay live after go-live. Any system that keeps writing data during the project turns a one-off migration into a continuing integration, which needs its own plan.
2. Profile the source
Count records per object, measure field fill rates and find the duplicates before anyone writes a mapping spreadsheet. Profiling routinely halves the volume people expected to migrate.
- Counts and fill rates. A field that is 4% populated is a candidate to drop, not to map.
- Duplicates. Match accounts on cleaned company name plus domain, and contacts on email. Decide the survivor rule before the load, not after.
- Orphans. Contacts with no account and deals with no company need a home or a decision.
- Picklist drift. Export the distinct values of every picklist-type field. "UK", "U.K." and "United Kingdom" become one value in the map.
- Owners. List every record owner, including people who have left.
3. Agree the field map
A field map is a contract. It names every source field, its destination, its transformation and what happens when the value is missing. Anything unmapped is a decision someone will otherwise make under pressure at 2am.
Keep one row per source field with these columns: source object and field, target object and field, data type, transformation rule, default value, and owner sign-off. Map inactive owners to a named holding user rather than dropping the records. Map every legacy primary key to an external ID field on the target object. That one column is what lets you upsert safely, rerun a failed batch and link child records without lookups by name.
4. Prepare the target org
- Create the external ID fields, record types and picklist values the map needs.
- Create users first, and decide which ones need licences on day one.
- If you must keep original created dates, enable "Set Audit Fields upon Record Creation" in Setup and grant it through a permission set to the integration user only. CreatedDate and CreatedById can be set on insert, not on update.
- Build a bypass for automation. A custom permission or custom setting that validation rules, flows and triggers check lets you skip them for the integration user during the load, without deactivating anything for real users.
- Set duplicate rules to allow or report for the integration user, so the load does not reject records you already de-duplicated.
5. Load in dependency order
Most "missing" data after a migration is a child record loaded before its parent. Load in this order and use external IDs for every relationship:
- Users and queues
- Accounts, then a second pass to set the parent account on hierarchies
- Contacts, then account-contact relationships if you use them
- Leads and campaigns, then campaign members
- Products, price books and price book entries
- Opportunities, then opportunity line items and contact roles
- Cases, then quotes, contracts, orders and custom objects
- Activities (tasks and events), then notes and files
For tools, the Data Import Wizard handles simple loads of up to 50,000 records on standard objects. Data Loader covers most SMB migrations up to a few million records. Above that, or where a load must run unattended, use the Bulk API 2.0 through Data Loader or an ETL tool. Whatever the tool, keep the success and error files from every batch.
6. Dry run in a sandbox
Load the full volume, not a sample. Volume exposes governor limits, timeouts and validation-rule collisions that a hundred-row test never will. Time the run. The timing tells you whether cutover fits in a weekend or needs a delta strategy, where you load history early and only the changes made since during the cutover window.
Run it at least twice. The first run finds the errors. The second proves the fixes and gives you a timing you can plan around.
7. Reconcile
Reconciliation is how you prove the data arrived, not just that the load finished.
- Row counts per object, source against target, with every difference explained.
- Sum totals for key numeric fields, such as opportunity amount by stage and invoice totals by year.
- Relationship checks: contacts without accounts, opportunities without accounts and line items without products should all be zero.
- Spot checks on the twenty most valuable records, opened side by side with the source by someone who knows those customers.
Sign the reconciliation report before cutover. For a typical SMB dataset it takes a day. It is never the step to compress.
8. Run the cutover
- Announce the source freeze time and switch users to read-only.
- Run the delta or full production load in the rehearsed order.
- Reconcile again, using the same report as the dry run.
- Turn automation, duplicate rules and integrations back on.
- Smoke test the top ten business processes with real users.
- Get the go or no-go decision from the person named in the plan.
9. Plan the rollback
Know what "abort" looks like: which loads reverse, which stay, who decides, and by what time. Because every record carries a legacy external ID, a rollback can delete exactly what the migration inserted and nothing else. Users keep working in the old system until the new one is signed off. A rollback plan you never use is the cheapest insurance in the project.
After go-live
Keep the source system read-only for at least 30 days, so anyone can check a value they doubt. Remove the audit-field permission and the automation bypass from the integration user. Watch the first month's duplicate reports, because new data entry exposes rules that the migration bypassed. If the move includes CPQ, our guide to migrating CPQ without freezing a quarter of selling covers the quote and pricing side.
Frequently asked questions
What should a Salesforce data migration plan include?
A scope list with a history cut-off, a profiling report, an approved field map, the load order, a sandbox dry-run timing, a reconciliation report template, a cutover runbook with named owners, and a written rollback plan.
Should we migrate all historical data?
Usually not. Migrate open records and the history people actually use, typically two to three years, and archive the rest. Dead data slows every future release and costs storage.
Which tool should we use: Data Import Wizard or Data Loader?
Use the Data Import Wizard for simple loads of up to 50,000 records on standard objects. Use Data Loader, or an ETL tool on the Bulk API, for anything larger, for custom objects at volume, or when you need repeatable upserts with external IDs.
Can we keep the original created dates?
Yes. Enable "Set Audit Fields upon Record Creation" and grant it to the integration user. Created and last-modified dates and users can then be set when records are inserted, not when they are updated.
How long does a Salesforce data migration take?
For a typical SMB moving a few hundred thousand records, two to four weeks from profiling to cutover, including two sandbox dry runs. Reconciliation takes about a day and should never be compressed.
Want this applied to your org?
CodeSierra runs fixed-scope Salesforce data migrations as part of every implementation, with the scope, timeline and price agreed in writing before the build starts. Ready to get started? Contact us for a scoped quote.