Data Warehouse Migration in 2026: When to Move and How to Cut Over Safely

Most data warehouse migrations are pitched as a choice between platforms. The record of companies that have done it says the platform matters less than three decisions made around it: how much of the old estate is retired before anything moves, how converted code is proven to give the same answers, and when the old system is finally switched off. This guide covers why 2026 forces the question, what the named migrations moved and gained, how far AI conversion tools can be trusted, and how to cut over without breaking the reports a business runs on.

• ~70%of LinkedIn's migration workload removed by consolidating 1,424 datasets into 450 before moving⁴
• End of 2027mainstream maintenance ends for SAP NetWeaver 7.5, the platform under SAP BW 7.5¹
• Under 38.5%average accuracy of language models translating SQL between database systems in an independent benchmark¹⁰

Why 2026 is a decision year

For many estates the vendor has already set the date. SAP's mainstream maintenance for NetWeaver 7.5, the platform SAP BW 7.5 runs on, ends at the end of 2027, with optional extended maintenance to the end of 2030 at a premium, while BW/4HANA follows S/4HANA's commitment to 2040.¹ Others are past their deadline and running on borrowed time.

AI adds a second reason. In a Gartner survey of 1,203 data management leaders, 63% said their organizations either lack or are unsure they have the right data management practices for AI, and Gartner predicts organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026.¹⁶ The prediction is a forecast, and its direction matches where the money is going. Snowflake's product revenue grew 37% year on year to $1.49 billion in its quarter to July 2026, and Databricks reported a revenue run-rate above $7 billion, growing more than 80%.¹³¹⁷ In a Dremio survey of 563 IT decision-makers, 55% already ran most of their analytics on a lakehouse and 67% expected to within three years.¹⁸ Dremio sells lakehouse software, so the direction is the useful part.

Open table formats moved the lock-in to the catalog

The biggest change since 2024 is where the data sits. Databricks bought Tabular, the company founded by Apache Iceberg's creators, to reconcile Iceberg with its own Delta Lake, conceding that the two formats "became incompatible due to their independent development."¹⁹ Snowflake made Iceberg tables generally available in June 2024 and its Open Catalog, previously named Polaris, in October 2024.⁵²⁰ AWS launched S3 Tables with built-in Iceberg support in December 2024, and Google made its BigLake Iceberg tables and metastore generally available in May 2025.²¹⁶ A table written once in customer-owned storage can now be read by several engines.

That shifts the lock-in instead of ending it. The Apache Polaris proposal itself says that "many interdependent limitations exist between engines and catalogs, which create lock-in."⁷ The two leading catalogs are both open source and compete: Databricks released Unity Catalog under Apache 2.0 in June 2024 with an Iceberg REST interface, and Apache Polaris became a top-level Apache project in February 2026.²²²³ Permissions, lineage and metric definitions live in that layer, so choosing the catalog deserves the care buyers used to give the database.

Open formats also make the next move cheaper, including a move off a cloud warehouse. New Relic moved more than 1,000 datasets from Snowflake to Iceberg and reports a 35 to 52% reduction in annual data platform spend, after finding itself "deeply locked into a single vendor's storage, compute, and query model."²⁴ Notion built its own lake in 2022 because 90% of its upserts were updates, a pattern most warehouses are not built for, and reports net savings of over a million dollars that year.²⁵

What the named migrations moved and gained

Companies publish what they moved and how fast it runs now. None of the companies below publishes the total cost of the project.

Three lessons repeat. Scope comes down before anything moves: LinkedIn used lineage to collapse its datasets, and PayPal switched off a quarter of its workloads.⁴²⁶ At very large scale, the data moves first and the code is modernized later, so users see no change on day one.²⁷ And running two warehouses has a price that grows with time. LinkedIn's first, unplanned attempt left both systems running with data copied between them, which it says "resulted in double the maintenance cost and complexity, along with very confused consumers."⁴

The failure statistic everyone quotes

The figure that 83% of data migrations fail or overrun is usually credited to Gartner, and we could not trace it to any Gartner publication. The traceable source is Bloor Research. Its 2007 survey found that 84% of data migration projects ran over time, over budget or both. Its 2011 survey found nearly 62% delivered on time and on budget, and 30% delayed, by about four months on average and sometimes a year or more.¹² Both surveys predate cloud warehouses, and no current independent survey replaces them.

The most expensive recent failure shows where the risk actually sits. When TSB moved its customer services to a new platform in April 2018, the Bank of England noted that the data itself migrated successfully. The platform then failed, all of TSB's branches and a large share of its 5.2 million customers were affected, and service did not return to normal until December 2018. Regulators fined TSB £48.65 million, on top of £32.7 million paid to customers in redress.²⁹ Moving the data is the part teams plan for. Proving the new system behaves the same under real load, and having a way back, is the part that fails.

AI conversion: fast on syntax, unproven on meaning

Every major platform now gives away its migration tooling, and each has added a language model on top of a rules engine. Snowflake made SnowConvert free in January 2025 and says it automates more than 96% of code and object conversion.⁸ Google says its Gemini-enhanced BigQuery translation delivers over 95% accuracy, Databricks says Lakebridge automates up to 80% of migration tasks, and AWS says its generative AI conversion handles up to 90% of schema objects for moves to PostgreSQL.³⁰⁹³¹ None of these figures comes with a published method, and each measures something different.

The independent numbers are lower because they check results. On PARROT, a benchmark of SQL translation across 22 database systems, language models averaged under 38.53% accuracy.¹⁰ On UniQL, the best model answered 54.63% of questions correctly across 16 dialects, scored lowest on Teradata at 37.74%, and got only 20.14% of questions right in every dialect.¹¹ Enterprise schemas are harder still. On ESQ-Bench, built on populated enterprise Oracle schemas, GPT-4o's execution accuracy fell from 79.8% to 57.2% as complexity rose, against more than 89% reported on academic benchmarks, and most failures at the harder tiers were queries that ran and returned wrong results.³²

The vendors' own documentation points the same way. Microsoft's Fabric Migration Assistant warns that "mistakes can happen as Copilot uses AI, so verify code suggestions before running them," and notes that T-SQL compatibility is incomplete.³³ AWS lists generative AI conversion as unavailable for Oracle to Redshift, its only warehouse target.³⁴ Snowflake's newer verification runs converted logic on both the source system and Snowflake with synthetic data and compares the results.³⁵ That is the right test for any tool: a converted query is accepted when it returns the same answer as the original on the same data.

Reconciliation is the acceptance test

Data checks run in tiers. Row counts by table and partition come first, then column aggregates, then row-level hash comparisons on the tables that feed money, risk and regulatory figures. Google's open-source Data Validation Tool covers all three levels across sources including Teradata, Oracle, Snowflake and BigQuery.³⁶ Datafold sunset its open-source data-diff package in May 2024 and kept diffing in its commercial product, a reminder to check the maintenance status of any free tool a migration depends on.³⁷

The acceptance threshold belongs to the business. In Faire's move of more than 5,000 tables from Redshift to Snowflake, each business group agreed a target level of accuracy with its data engineers before signing off, according to a case study published by the diffing vendor.³⁸ Databricks' migration team runs parallel testing and reconciliation with business experts in user acceptance testing, starting from a production pilot of one end-to-end use case.³⁹ Agreeing tolerances before the first comparison keeps sign-off from turning into a negotiation at the end.

The cheapest table to reconcile is the one that never moves. In a migration Microsoft documents, a consumer goods company found that about half of its thousands of reports had been opened in the previous year, and about half of those delivered significant value. Microsoft's advice on the rest is blunt: "Sometimes, the cheapest and easiest thing to do is nothing."⁴⁰ Metric definitions deserve the opposite treatment. They drift quietly when SQL is rewritten, so writing each one down once, in a portable form such as the open-source Open Semantic Interchange specification or dbt's MetricFlow, makes them testable and keeps them out of the next lock-in.⁴¹

How the bill behaves after go-live

Consumption platforms charge for capacity running, and the idle traps differ by platform. Snowflake bills per second with a 60-second minimum each time a warehouse starts, and its resource monitors cover warehouses only, leaving serverless and AI features to separate budgets.⁴²⁴³ BigQuery charges for the slots it scales up to, not the slots a job uses, even if the job fails, and keeps scaled capacity for at least 60 seconds.⁴⁴ Redshift Serverless keeps billing for open transactions, cancelled queries and the health-check queries connection pools send.⁴⁵ Databricks says its budgets should not be used "as a way to ensure an absolute spend cap."¹⁴

Growth after adoption is the design. Snowflake's net revenue retention of 126% means its existing customers spent 26% more than a year earlier.¹³ Instacart's 2023 filing showed payments to Snowflake of about $13 million, $28 million and $51 million from 2020 to 2022; Snowflake replied that actual usage was $28 million in both 2021 and 2022, and that Instacart had optimized its Snowflake costs by 50%.⁴⁶⁴⁷ Prepaid commitments and real consumption can tell very different stories, and tuning moves the bill as much as the platform does. Flexera's 2026 respondents put wasted cloud spend at 29%, the first rise in five years, and the FinOps Foundation finds data cloud platforms and AI are now the most actively managed areas because of "less predictable usage patterns."⁴⁸⁴⁹

In regulated firms, the migration is a supervised change

The ECB lists reconciliation errors among the causes of large-scale miscalculations of key risk ratios it has found at banks.⁵⁰ DORA's exit requirement also turns this migration into a design brief for the next one: open formats, a catalog that other engines can read, and code kept in a repository make a tested exit possible.⁵¹

How to run a data warehouse migration

1. 1Inventory usage and retire first Read the audit logs and lineage of the old platform, list who uses each report and table, and retire what nobody opened in a year. Every object retired is one fewer to convert, test and reconcile.
2. 2Choose the format and catalog before the engine Keep tables in an open format such as Iceberg in storage you own, and pick the catalog for permissions, lineage and the engines that must read it.
3. 3Write the metrics down once Define revenue, margin, active customer and every other headline number in one portable semantic layer before any SQL is rewritten, so old and new can be compared on the same definitions.
4. 4Convert with tools, accept on results Let the rules engine and the language model do the first pass, then accept code only when the converted version returns the same results as the original on the same data.
5. 5Reconcile in a parallel run Run counts, aggregates and row-level hashes on every load cycle, against tolerances the business owner agreed in advance, until each domain has a run of clean cycles.
6. 6Cut over by domain and switch off on a date Move one domain at a time with a tested way back, and set the date the old platform stops for that domain before the work starts, so two systems never run indefinitely.
7. 7Turn on spend controls on the first day Set warehouse monitors, maximum capacity limits and budgets per team before users arrive, and review the first month's bill line by line.

The platforms will keep converging, and the free tools will keep improving at the parts that are easy to measure. What separates a migration that pays from one that drags on is still the unglamorous work: an estate made smaller before it moves, numbers that match to an agreed tolerance, and an old system that is actually switched off.

This is how we run data engineering work: an audit of what is used and what each number means, open formats in your own cloud account, reconciliation built into the pipeline, and the old platform retired source by source. On one hospital ERP, putting reporting on a single layer took administrative reports from days to minutes.