TL;DR — The table-format war was never about file layouts; it was about gravity. Databricks open-sourced Delta Lake in 2019 and finished the job with Delta 2.0 in June 2022, while Apache Iceberg — born at Netflix, donated to the ASF, backed by everyone who wasn't Databricks — became the neutral standard. In June 2024 Snowflake launched Polaris Catalog and Databricks answered by buying Tabular, the company founded by Iceberg's own creators, announced during Snowflake's conference week.
Yesterday's benchmark war fizzled without a referee — two vendors trading TPC-DS numbers until one of them declared the whole exercise stupid and walked. The format war was the opposite. It had a clear resolution, it took five years, and it ended with one company writing a very large cheque for the other side's founders.
Here is why it mattered enough to be worth that cheque. A table format is boring plumbing that decides which engines can read your data, and whoever controls it controls the gravity: if your petabytes are only fully legible to one vendor's compute, that vendor's compute is the one you buy next year. Both companies understood this. Both spent the decade making sure the standard was either theirs or nobody's.
The concepts you need
Five ideas carry this story.
Concept: an open table format. Object storage holds Parquet files and has no idea they're a "table." A table format is metadata beside the data saying these files are the table right now. It buys four things. Atomic commits: the new pointer publishes only once the whole batch succeeded, so readers never see half a job. Snapshots: every commit makes an immutable version, valid until expired. Time travel: query the table as it was last month — also how you undo a bad job. Schema evolution: renaming a column edits metadata, not petabytes, because columns carry stable IDs, not positions. All three formats do this. They differ on layout, and on who decides.
Concept: a catalog. The format tells an engine how to read one table; the catalog tells it which tables exist, where they live, and whether you may touch them. It's the namespace, the referee when two engines commit to the same table, and the source of short-lived scoped credentials instead of your bucket keys. The Iceberg REST spec standardised that as an HTTP API any engine can speak — which is why the catalog became the new control point.
Concept: Apache governance vs vendor-controlled open source. Both say "open source"; the risk differs. In an ASF top-level project the spec is foundation-owned and changes go to a public vote, so no company can hold a feature back. Vendor-controlled means a permissive licence but roadmap and release timing inside one company, free to ship to the product first and the repo later. The test isn't "is there a licence," it's "can one company withhold something and still call it open?"
Concept: an acqui-hire of a standard. You can't buy an Apache project — there's no cap table. You can buy the company employing the people whose commits and votes steer the spec — influence rather than assets, which is how a price tag detaches from revenue.
Concept: interoperability by dual metadata (UniForm). The formats share the same Parquet files, so a Delta table can also emit Iceberg- and Hudi-shaped metadata pointing at those untouched files. An Iceberg-only engine reads it natively and never knows: one dataset, two passports, no copy. The catch: the second passport is generated, so it can lag the spec it imitates.
The three contenders
All three sit on the same Parquet data files. The differences are governance and metadata — politics and physics.
| Delta Lake | Apache Iceberg | Apache Hudi | |
|---|---|---|---|
| Born at | Databricks | Netflix (~2017, per press reports, created by Ryan Blue and Dan Weeks to fix Hive's correctness and performance problems) | Uber, starting 2016 |
| Open-sourced | April 24, 2019, at Spark+AI Summit, announced by CEO Ali Ghodsi | Donated to the Apache Software Foundation November 2018 (reported) | Open-sourced 2017 |
| Governance | Linux Foundation from October 2019 | Apache Software Foundation | Apache Software Foundation |
| Top-level status | n/a (LF project, not ASF) | Graduated to Apache top-level May 20, 2020 (reported) | Entered the Apache Incubator January 2019; graduated top-level June 2021 |
| The "fully open" moment | Delta Lake 2.0, announced June 28, 2022 at Data+AI Summit — contributed all previously-proprietary APIs and features to the open project | Open from donation; no single-vendor holdback to undo | Open from donation |
| Who backs it | Databricks, plus the Linux Foundation ecosystem | Effectively everyone else: AWS, Google Cloud, Microsoft, Confluent, Dremio, Salesforce all show up as backers of Iceberg-native catalog efforts | Uber-origin community; narrower vendor coalition |
| Cross-format bridge | Delta UniForm (Delta Lake 3.0, June 28, 2023) writes Iceberg- and Hudi-compatible metadata alongside Delta | Is the thing others bridge to | Readable via UniForm; also a UniForm target |
| Status 2025–26 | Persists as Databricks' native format; Databricks added native Iceberg support June 2025 (reported) | The cross-vendor interoperability standard; v3 spec ratified 2025 (reported) | Alive for differentiated use cases; hit a 1.0 milestone, alongside newer entrants like Apache Paimon |
Read that "fully open" row twice: it is the whole argument of 2019–2022 in one line. Delta was open source from 2019, but not all of it — some features stayed proprietary until 2.0, which is what "open-but-Databricks-controlled" meant, and it was structurally true, not a competitor talking point. Delta 2.0 closed the gap. But by June 2022 Iceberg had been an Apache top-level project for two years, with no vendor to hold anything back in the first place. Iceberg didn't win on features. It won on the absence of a landlord.
June 2024: the week the war ended
The endgame ran nine days, and the calendar is the story.
| Date | Move |
|---|---|
| June 3, 2024 | Snowflake launches Polaris Catalog at Snowflake Summit — an open, vendor-neutral Apache Iceberg REST catalog. Launch backers: AWS, Google Cloud, Microsoft Azure, Confluent, Dremio, Salesforce. |
| June 4, 2024 | Databricks announces it is acquiring Tabular — the company founded by Iceberg's original Netflix co-creators — during Snowflake's own conference week. Snowflake stock falls roughly 7%. |
| June 12, 2024 | Databricks open-sources Unity Catalog, with a native Iceberg REST catalog and Hive metastore API compatibility. Backers: AWS, Google Cloud, Microsoft, NVIDIA, Salesforce. Over 10,000 enterprises were reported on managed Unity Catalog at announcement. |
Snowflake's play was coherent: if Iceberg is the neutral format, be the neutral catalog on top of it. Polaris was open-sourced under Apache 2.0 within about 90 days, donated to the ASF around August 2024, shipped a 1.0 in October 2025 with external IdP support, and — per reporting — graduated to full Apache top-level project status on February 18, 2026. That is a company giving away the control point on purpose, because being the default beats being the owner.
Databricks' counter was blunter. Ghodsi confirmed the Tabular deal came out of a bidding war that included Snowflake and Confluent. The timing — landing the announcement inside a rival's conference — was read exactly as intended.
On the price, be careful. Nobody has confirmed it. Initial June 2024 reporting put it near $1 billion, largely performance- and earn-out-based, with Ghodsi indicating near-term cash outlay was likely "in the hundreds of millions." Bloomberg reporting relayed by TechCrunch in August 2024 put it closer to $2 billion. Databricks has never disclosed the figure. Tabular had roughly 40 employees, had raised over $30 million, and was reported to have around $1 million in annual recurring revenue — three years after founding. Whatever the number was, it wasn't for the revenue.
What Databricks actually bought
Not a product. Iceberg is Apache-governed; you cannot buy a format. What was purchasable was the people whose commits and design authority shape where the format goes, plus the credibility of saying "the folks who created Iceberg work here now" while your competitor stakes its interoperability story on Iceberg.
With UniForm, Databricks' position becomes a hedge rather than a bet. Delta stays native; Iceberg gets read and written natively as of June 2025 (reported). If the market picks Iceberg, Databricks employs its creators. If Delta holds, Databricks owns it. There is no branch of that tree where Databricks loses the format.
Where convergence stands, 2025–2026
Per the pack's reporting, the war has substantially converged on Iceberg as the interoperability standard. The v3 spec was ratified in 2025 — binary deletion vectors, a variant type for semi-structured data, row lineage, geospatial types, nanosecond timestamps. AWS announced v3 deletion-vector and row-lineage support in November 2025. Google BigQuery enabled Iceberg-compatible read/write interop in April 2026. Every major cloud ships a managed Iceberg service.
The battleground moved up the stack. It is now catalogs — Polaris versus Unity Catalog versus the raw Iceberg REST spec. A better fight to be having, but note what it is: the same gravity argument, one layer higher.
Delta stack vs Iceberg stack: the honest trade-offs
Neither choice is wrong in 2026, and neither is free.
| Delta stack | Iceberg stack | |
|---|---|---|
| Advantages | Native to the Databricks runtime, so integration is deep by construction and features land there first. Grew up in Spark — the most exercised code path in the category. One dominant implementation means fewer "which flavour of the spec does this engine speak" surprises. Since Delta 2.0 the licence question is settled, and UniForm gives a metadata-level exit with no data copy. | Apache top-level since May 20, 2020 (reported), so no vendor can withhold a feature and still ship the spec. Broadest engine support anywhere — the REST spec is spoken by Spark, Flink, Trino, PyIceberg, StarRocks and Doris, every major cloud ships a managed service, and both combatants read and write it natively. |
| Disadvantages | The governance history: open-sourced April 2019, features proprietary until 2.0 three years later. Parts of the ecosystem still price in vendor risk the current licence no longer justifies. Stewardship is Linux Foundation, not ASF, and roadmap authority isn't meaningfully distributed. Outside Spark and Databricks, support is thinner and often arrives via UniForm translation rather than a native reader. | The catalog story is fragmented in a way the format story isn't: Polaris, Unity Catalog, raw Iceberg REST and each cloud's managed service are all legitimate answers, and choosing carries switching cost. "Supports Iceberg" spans read-only to read/write and one spec version to another, and says nothing about who owns the maintenance (compaction, snapshot expiry) that keeps tables fast. |
And the catalog is where lock-in now lives. The open catalog and the managed catalog are not the same product. Policies, credential vending and lineage accumulate in the managed one, and none of that is described by the Iceberg spec. Storage is portable; the permission model you spent two years building is not.
The developer takeaway. Storage format is no longer a lock-in decision worth agonising over. Pick Iceberg unless you have a specific reason not to; if you're already on Delta, UniForm makes the exit cheaper than a migration. Spend the paranoia budget on the catalog and the governance model instead. Ask a prospective vendor which catalog spec they implement, not which format they write.
So which company is actually more open source?
Since this is the open-source day, let's answer the question directly, because the two companies are not close on this axis — and the honest answer has a twist at the end.
Databricks is open at the core. The engine it was founded on, Apache Spark, is an ASF top-level project anyone can run without paying Databricks a cent. Its table format, Delta Lake, has been fully open since Delta 2.0 in June 2022 (Linux Foundation). MLflow, the experiment-tracking standard, is theirs and open. Unity Catalog got an OSS release in June 2024, DBRX shipped open weights in March 2024, and the Neon acquisition brought in an open-source serverless Postgres. The closed parts are the money parts: Photon — the C++ engine that actually wins the benchmarks — is proprietary, and the managed platform is the product.
Snowflake is open at the edges. The core engine has been deliberately closed since 2012 — that was the founding bet, and it has never changed. Its open-source moves are real but peripheral: Streamlit (open app framework, acquired 2022), Arctic's open weights under Apache 2.0 (April 2024), Snowflake Postgres via Crunchy (June 2025), and Polaris (June 2024). Snowpark, despite the developer-friendly branding, is a client API into a closed engine, not an open engine.
So by volume and by centrality, Databricks is the more open-source-based platform, and it isn't close — you can assemble most of its stack from public repos, while you cannot run one line of Snowflake's engine yourself. But here is the twist worth keeping: the single cleanest open-source governance act of this whole rivalry came from Snowflake, which donated Polaris to the Apache Software Foundation outright — surrendering control in a way Databricks has never quite matched, with Delta and Unity Catalog living in foundations where Databricks retains effective roadmap authority. Databricks is more open; Snowflake, on the one occasion it opened something structural, let go more completely. Both facts belong in your evaluation.
What to Learn From This
- Evaluate governance, not the licence badge. Does a foundation own the spec, is there more than one serious implementation, and do features hit the public repo the same day as the product?
- Track where lock-in moved, because it always moves up. Commoditise a layer and the fight relocates one higher — format then, catalog now. Ask who vends credentials and whether your permission model is exportable.
- Design your exit as a metadata change, not a migration. UniForm works because all three formats share the same Parquet files. Generalise it: open files underneath, thin engine-specific layer on top.
- Standards are made by people, and people can be hired. Design authority is purchasable even when the spec isn't. Read the committer list and their employers, not the foundation logo.
- Test multi-engine support at the seams. Verify spec version, read versus write, who runs compaction and snapshot expiry, and what happens on concurrent writers — the last one is where cross-engine setups actually break.
Scorecard
Round 5: Databricks. On execution, this isn't close. Databricks open-sourced its format, then closed the credibility gap with Delta 2.0, then built UniForm so its format could impersonate the competition's, then bought the competition's creators and announced it during the competition's conference — reportedly outbidding Snowflake for them. It ended the round holding both formats and a catalog with 10,000+ enterprises on the managed version. Snowflake played it well and arguably more honourably, giving Polaris away to the ASF outright, but it spent the week being the backdrop to someone else's press release.
The deeper truth is that neither vendor won the war. Apache Iceberg did — a format built at Netflix, governed by a foundation, adopted by every cloud, and now supported natively by both combatants. Databricks takes the round on tactics; the open standard took the decade.
Running tally: Databricks 2 — Snowflake 1 (2 draws)
Tomorrow
Part 6: the AI pivot — a $1.3B model shop, a founder-CEO swap, and the agent race.
Sources
- Databricks launches Delta Lake, an open source data lake reliability project (VentureBeat, Apr 2019)
- Delta 2.0 — The Foundation of your Data Lakehouse is Open (Linux Foundation, 2022)
- Announcing Delta Lake 3.0 with Universal Format (PR Newswire, Jun 2023)
- Snowflake unveils Polaris Catalog (Business Wire, Jun 3 2024)
- Databricks is buying data optimization startup Tabular (CNBC, Jun 4 2024)
- Databricks reportedly paid $2 billion in Tabular acquisition (TechCrunch, Aug 2024)
- Open sourcing Unity Catalog (Databricks, Jun 12 2024)
- What's new in Apache Iceberg v3 (Google Open Source Blog, Aug 2025)
Top comments (0)