Nobody plans to hit an iceberg. The Titanic's officers had charts, lookouts, and wireless warnings from three other ships, and still misjudged what lay below the surface. Data platforms fail for a quieter version of the same reason. A company picks a storage format because it shipped fast or because a vendor demo looked clean, and five years later the real cost surfaces: the part of the system nobody budgeted for.

That question, what's actually underneath, is what tends to send companies looking for an outside opinion in the first place. A firm offering data lake consulting generally isn't hired to explain what a data lake is; it's hired to prevent the 5-year surprise. Guidance on structuring a data lake has become one of the more common engagements a business starts once the decision starts to feel irreversible. Ask 3 different consultants which table format to pick this year, and there's a decent chance all 3 land on the same name: Apache Iceberg.

Three Formats, Three Very Different Boats

Apache Iceberg started as an internal fix at Netflix, built to stop queries from quietly returning wrong answers because of how Hive tracked partitions. Ryan Blue's team donated the specification to the Apache Software Foundation in 2018, and that detail matters more than it sounds like it should. No single company owns the roadmap. Snowflake, BigQuery, Databricks, Trino, and DuckDB can all read and write native Iceberg tables today, so a business isn't locked to whichever vendor sold the platform. Portability isn't a footnote here.

Delta Lake tells a different story. It grew up inside Databricks, wound tightly around Apache Spark, and for teams already living in that world, the fit stays close to frictionless. Microsoft Fabric ships Delta as its default table format, and Databricks itself still claims the deepest installed base of the three, touching a large share of Fortune 500 companies through one product line or another. None of that makes Delta the wrong choice. It just makes it a choice with a center of gravity, one that pulls toward a single vendor's tooling even when the documentation insists otherwise.

Apache Hudi solves a narrower problem, and solves it well. Built at Uber to handle constant, high-speed upserts, it targets workloads where yesterday's data is already stale by lunchtime. Companies moving change-data-capture streams still lean on Hudi's write path, since little else in the category handles record-level updates at that pace as cleanly. A table format, in the end, is closer to plumbing than to a dashboard. Nobody notices it until it leaks.

Which Way the Current Is Pulling

Numbers help settle arguments opinions can't. Analysts at Mordor Intelligence project the data lakes segment climbing from close to $18.7 billion in 2025 to about $22.8 billion this year, with growth still holding near 22% annually through 2031.

A fair share of that expansion traces back to companies rebuilding around AI pipelines that need clean, queryable data at a scale spreadsheets were never built for. Cloud vendors noticed early. Built directly into the storage layer, Amazon's S3 Tables ship as the first cloud object store with built-in Apache Iceberg support, compatible out of the box with engines ranging from Spark and Trino to Snowflake and Redshift. Google and Microsoft have made comparable moves on their own storage layers. When all three hyperscalers point the same direction, that's not a trend piece. That's a market making up its mind. Buyers appear to be moving with it: research covered by HPCwire's BigDATAwire found that among more than 560 IT decision-makers surveyed, the share planning to run most of their analytics on a lakehouse architecture within 3 years jumped past two-thirds, up from just over half the year prior. For its part, Databricks paid over a billion dollars in 2024 for Tabular, the company Iceberg's original creators founded after leaving Netflix, largely to keep pace with a format its own product line didn't invent. Hudi answered on its own terms. Its January 2025 release added the ability to output native Iceberg-compatible metadata, letting teams keep Hudi's write-path tooling while producing tables any Iceberg reader can open.

For a business weighing all three, the shorthand outside advisors on data lake strategy tend to reach for looks something like this:

• Apache Iceberg fits the business running more than one query engine, more than one cloud, or both, and wants the freedom to swap either later without rewriting the data.
• Delta Lake makes the most sense inside a Databricks-heavy or Fabric-heavy shop, where deep integration outweighs format neutrality.
• For workloads built on constant, record-level updates, streaming ingestion and CDC pipelines especially, Hudi still carries the most mature tooling of the three.

None of the three is wrong, and picking one that fights the actual shape of the workload is.

The Question a Good Consultant Actually Asks

Which format is best rarely turns out to be the right question. Best for what is closer to the real question, tied to an engine already running in production and a downstream report that finance depends on. A retailer syncing inventory across four systems in real time needs something different from a media company running nightly batch reports for finance, even if both call the end result a data lake.

Firms like N-iX have built dedicated data lake consulting practices partly because this fit question doesn't resolve itself. It takes someone outside the immediate team, someone without a stake in which vendor gets the invoice, to lay the trade-offs out plainly. That's less about naming a winner and more about mapping which format survives the next three platform decisions the business hasn't made yet.

The interoperability tools help here too. Delta's Universal Format and the still-young Apache XTable project both let a single physical table get read as more than one format, which lowers the cost of getting the initial choice slightly wrong. Nobody has to bet the company on a decision made in one afternoon anymore.

Conclusion

Apache Iceberg isn't winning because it dazzles. It's winning because it asks the least of a business later, once the platform changes, the vendor changes, or the team does. That's a quieter kind of victory, and probably the only kind worth trusting in infrastructure. Whether it fits a given data lake consulting engagement still depends on workloads no market report can see from the outside, which is exactly the gap firms like N-iX are built to close. Best, more often than not, begins with someone asking what's really sitting below the