· Updated · 13 min read
Migrating SQL Server and Oracle to Snowflake: Moving the Data and Validating the Cutover
A defensible cutover joins the bulk load, parallel sync, parity evidence, and rollback conditions, so you can decide when Snowflake can replace the source.
Loading the rows alone does not make a cutover defensible. You must also manage downtime, cost, operational risk, and final approval. A migration is not a copy job in which a successful load proves that Snowflake can replace the source. Switch off SQL Server or Oracle only when Snowflake meets the agreed acceptance criteria.
What is a data warehouse migration?
A data warehouse migration moves the data downstream users need to a new warehouse and proves the new warehouse can replace the old one. For Snowflake, this includes the historical load, synchronization of later changes, type decisions, validation, cutover, rollback, and monitored use.
Define the acceptance criteria and rollback conditions before selecting an option. The criteria specify which changes must reach Snowflake, which comparisons must pass, and which conditions halt the cutover.
Required quality depends on how people use the data. The UK framework calls this “fitness for purpose.” It makes user needs central and requires assessment throughout the data lifecycle.
How do you load historical data and keep Snowflake in sync?
Plan the historical load and ongoing synchronization together. A bulk copy cannot support a low-downtime cutover unless changes made during and after the copy also reach Snowflake. Stripe’s migration combined a historical backfill with subsequent changes before comparing production results.
A long extraction also needs a stable point in time. SQL Server can hold one snapshot transaction across chunks, but retained row versions can increase storage pressure on the source. Oracle can apply one System Change Number across its chunks. If Oracle loses the required undo, the extraction can fail because the earlier version is no longer available.
Choose the synchronization mechanism from the changes that the cutover must preserve. Change Tracking records keys, versions, and operations, but collapses multiple changes. SQL Server Change Data Capture records captured values and update images. Oracle redo mining can expose deletes, transaction order, rollback state, and DDL through LogMiner records. If new and updated rows are enough, an incremental query can meet the acceptance plan. Use log-based capture when you require deletes, transaction order, intermediate changes, or lower latency. If OpenFlow misses the SQL Server Change Tracking retention period, the affected tables require a full reload. Lost Oracle undo can instead make the extraction fail. For broader method details, see the Snowflake ingestion guide.
Which SQL Server and Oracle data types need a decision?
Define acceptable data type mapping before loading. Start with the meaning that downstream users need, then verify how your selected product handles the source type. A conversion that can alter that meaning needs an explicit decision before you set validation thresholds.
The table identifies the main decisions and supported checks. It does not define a universal Snowflake mapping.
| Source | Type or risk | Decision before loading | Supported check |
|---|---|---|---|
| SQL Server | Decimal precision and scale | Preserve, round or use text | Compare numeric summaries |
| SQL Server | Date and time values | Preserve required time meaning | Use one numeric representation |
| SQL Server | Large text values | Set acceptable target size | Compare string lengths |
| SQL Server | Unique identifiers | Approve text representation | Compare sampled records |
| SQL Server | JSON or XML | Approve target representation | Compare sampled records |
| SQL Server | Geography or geometry | Choose required representation | Compare sampled records |
| Oracle | Unbounded NUMBER values | Set precision or use text | Compare numeric summaries |
| Oracle | Timestamp values | Preserve required time meaning | Use one numeric representation |
| Oracle | Interval values | Approve text representation | Compare sampled records |
| Oracle | CLOB and NCLOB values | Set acceptable target size | Compare string lengths |
| Oracle | JSON or XML | Approve target representation | Compare sampled records |
| Oracle | Other unlisted types | Approve text representation | Compare sampled records |
Numeric rounding can alter maxima, minima, averages, and other numeric summaries. Convert source and target date or time values to the same numeric representation before comparison. Text truncation appears through string-length differences. If a summary fails, use a field comparison to locate the affected record. Evidence did not establish a specific time zone or binary-value test.
What proves Snowflake is ready, and how long should both systems run in parallel?
Parity means that Snowflake is fit for the agreed downstream use. One row count cannot establish this result. Define the evidence, threshold, owner, and failure response before the parallel run starts.
Use several forms of evidence because each answers a different question. Each threshold belongs to your approved acceptance criteria.
| Evidence | What it establishes | Limitation | Threshold | Owner | Response |
|---|---|---|---|---|---|
| Row counts | Expected records arrived | Values can differ | Project-specific | Named in tracker | Correct and rerun |
| Per-table aggregates | Calculated property agrees | Offsetting errors can hide | Project-specific | Named in tracker | Isolate affected data |
| Hashes | Represented values agree | Representation must match | Project-specific | Named in tracker | Recompute after correction |
| Sampled rows | Tested records agree | Cannot prove full equality | Approved sample rule | Named in tracker | Draw a fresh sample |
| Downstream outputs | Required results agree | Covers selected uses | Approved report tolerance | Named in tracker | Track and correct defect |
Equal counts can hide offsetting omissions and duplicates, so record counts do not prove value equality. An aggregate establishes only its calculated property. Hashes require consistent parsing, canonicalization, and byte conversion. Statistical samples provide evidence at a stated confidence level, not full equality. Isolate failures by table, chunk, or record, correct them, and repeat validation.
Technical copy checks do not prove that reports and business results agree. Treat business validation as separate work. Record each discrepancy as a defect, assign an owner, and rerun failed work after correction. End the parallel run when these checks pass over the business periods that the migration must represent. There is no universal number of days. If a comparison fails, delay the switch.
Source: Stripe migration
How should you cut over, roll back and monitor after cutover?
Make a go/no-go cutover decision based on documented comparison results and open defects. Keep each threshold and trigger specific to the project.
| Condition | Evidence | Owner | Action | Rollback trigger |
|---|---|---|---|---|
| Checks pass | Approved results | Named owner | Approve cutover | Later critical mismatch |
| Checks fail | Recorded discrepancy | Named owner | Correct and retest | No cutover |
| Critical defect remains | Open defect record | Business owner | Delay cutover | No cutover |
| New system fails acceptance | Monitored result | Named owner | Restore prior service | Approved project trigger |
Keep the old store while dependencies remain and while the project still requires it for recovery. Stripe retained its old representation until no code depended on it. Use operational impact to set recovery priorities. NIST guidance connects tolerable downtime, recovery objectives, resources, and activity order, but it does not prescribe migration thresholds.
After cutover, repeat the approved report and business-result comparisons. Track open validation defects and their owners. Test and update recovery procedures. If you use Openflow, also review its Snowflake consumption.
Which option fits your migration?
Choose the option that preserves the required changes and meets the acceptance plan. Then compare Snowflake consumption, license cost, team time, downtime, operational risk, and lock-in.
The three options handle historical data and later changes differently.
| Option | Best fit | Change mechanism | Type handling | Snowflake consumption | License cost | Team time | Downtime | Risk | Lock-in |
|---|---|---|---|---|---|---|---|---|---|
Snowflake COPY INTO; Openflow | Exported files and managed replication | SQL Server tracking; Oracle XStream | SQL Server mappings; Oracle mappings | Warehouses plus Openflow | BYOC: 0.0225 credits • per vCPU-hour | Configure sources and flows | Set by cutover plan | SQL Server expiry can force reload | Definitions export; state export undocumented |
| Fivetran Enterprise | Managed database replication | SQL Server methods; Oracle methods | Unsupported types can change or skip | Snowflake compute billed separately | Monthly Active Rows | Configure and remediate connectors | Set by cutover plan | Method-specific source limits | Connector schemas and modes vary |
| dltHub | Teams that own pipeline code | SQL Server tracking; incremental queries | Reflected types with type overrides | Staged loading to Snowflake; Snowflake separate | USD 12,000 monthly | Team owns pipeline and checks | Set by cutover plan | Team preserves pipeline state | Repository and code retained |
Keep the cost components separate. Openflow BYOC charges 0.0225 Snowflake credits per vCPU-hour as of 30/10/26. Fivetran Enterprise bills by Monthly Active Rows, while Snowflake remains separate as of 30/10/26. dltHub costs USD 12,000 monthly, with a 12-month minimum and 5,000 monthly runtime credits as of 30/10/26. Snowflake remains separate.
Operation also changes the workload. Snowflake operates hosted OpenFlow infrastructure, while customers use its monitoring and diagnostics. Fivetran operates managed connectors, while customers follow the displayed remediation steps. OpenFlow flow definitions are exportable, but runtime-state export is not documented. Destructive removal can lose local state. Connector schemas and supported modes vary by connector.
Pricing and product information come from each vendor's public website and documentation, linked in the text, as reviewed on 30/10/26. Vendors change prices and plans without notice, so check the linked source or contact the vendor for current terms. Where a vendor does not publish a price, the article says so.
How can a team use dltHub to move the data and run its checks?
dltHub fits SQL Server migrations that can use Change Tracking. It also fits Oracle migrations when incremental queries for inserts and updates are enough. The sql_database source selects SQL Server or Oracle tables and views. You can divide historical loads into bounded ranges or ordered chunks. The team selects the tables, ranges, ordering, source-load controls, and type overrides. Later runs retain incremental state, while the SQL Server option can use Change Tracking with merge behavior and delete markers.
Reflected types become dlt types. The Snowflake destination stages files and loads them with COPY INTO. Its merge behavior needs keys for matching, while replace provides several table-replacement strategies.
Data quality checks, currently in preview, as of 30/10/26, can run after loading. Metrics apply at the column, table, or dataset level; checks apply at the row, table, or dataset level. is_unique and is_not_null are checks, while row_count is a table metric. Source-to-Snowflake row and aggregate comparisons remain Python jobs that the team writes and runs after loading. The managed runtime deploys, schedules, runs, and monitors the pipeline and adjacent jobs. Your team still chooses credentials, schedules, acceptance thresholds, comparisons, failure responses, and cutover or rollback approval.
Frequently asked questions
Do you need CDC, or are incremental loads enough during the parallel run?
Use log-based change data capture when you need deletes, transaction order, intermediate changes, or lower latency. SQL Server Change Data Capture and Oracle redo expose the relevant change details. Incremental queries can cover new and updated rows when those requirements do not apply. The Snowflake ingestion guide linked earlier explains the broader ingestion methods.
Who handles SQL dialects, stored procedures and ETL jobs?
This guide covers data movement and cutover validation, not code conversion. SnowConvert AI covers SQL Server and Oracle SQL and stored procedures. Snowflake’s migration partners can assess the work, correct converted output, test it and support implementation.
Book a call to assess dltHub for SQL Server Change Tracking or Oracle incremental loads where inserts and updates are enough.
What is dltHub?
The agentic data layer on top of any data warehouse. Build, govern, heal and serve your data infrastructure on dltHub's managed infrastructure.
Was this page helpful?