▣ dataarchitect.studio

Field Notes

Iceberg vs Delta Lake: How to Actually Choose in 2026

July 6, 2026/ Updated August 8, 2026/ 6 min read/Architecture

For years, choosing between Apache Iceberg and Delta Lake was the defining architecture decision of a lakehouse. In 2026, it mostly isn’t. The two open table formats have converged on capabilities — ACID commits, schema evolution, time travel, row-level deletes via deletion vectors — and interop layers can present one format as the other. The honest decision rule now fits in two sentences: inside Databricks or Microsoft Fabric, use Delta — it’s the native format and everything is tuned for it. Everywhere else, or for anything multi-engine, default to Iceberg — it has become the neutral standard that every major platform reads and writes.

Here’s the reasoning, and the one decision that still deserves your attention more than the format.

Iceberg vs Delta Lake, feature by feature

  Apache Iceberg Delta Lake
Governance Apache Software Foundation, multi-vendor Linux Foundation, Databricks-led
Core guarantees ACID, schema evolution, time travel Same
Row-level deletes Deletion vectors (v3) Deletion vectors
Semi-structured Variant type (v3) Variant type
Native home Snowflake, BigQuery, AWS, Trino, Flink Databricks, Microsoft Fabric
Multi-engine story Native — REST catalog spec Delta Kernel (4.0), UniForm
Interop escape hatch Read/written by nearly everything UniForm presents Delta as Iceberg
Choose it when Open, multi-engine, vendor-neutral Databricks/Fabric-centred platform

The capability rows are the point: they’re nearly identical. Both store data as Parquet. The convergence took different routes and is older than most comparison posts admit: Delta added deletion vectors incrementally from 2.3.0 in 2023, reaching full MERGE support and on-by-default behaviour in 3.1.0, while Iceberg standardised them alongside the variant type in the v3 spec. Performance differences are workload- and engine-specific, not categorical. Anyone selling you a decisive feature gap is reading from an old slide.

What actually differs: gravity

Delta’s gravity is Databricks. It’s the native tongue of the platform where a majority of Fortune-500 data teams already run, and Fabric adopted it too. Inside that world, Delta is simply less friction — and UniForm lets a Delta table publish Iceberg metadata so outside engines can still read it.

Iceberg’s gravity is everyone else. Snowflake ships first-class Iceberg tables, BigQuery reads it through BigLake, AWS backs it across Glue and Athena, and the Iceberg REST catalog spec is what engines standardise against. Even Databricks now writes Iceberg through Unity Catalog. When the whole industry needs one table language to interoperate, it picked Iceberg.

Choosing between Apache Iceberg and Delta Lake A decision tree beginning with where compute lives now and in three years. Mostly Databricks or Fabric branches to Delta Lake, with UniForm for outside readers. Mixed, multi-engine or uncertain branches to Apache Iceberg as the neutral default. Both branches converge on a final box: choose the catalog deliberately, because that is where lock-in now lives. Where does your compute live, today and in 3 years? mostly Databricks / Fabric mixed / multi-engine / unsure Delta Lake + UniForm for outside readers Apache Iceberg the neutral default either way: choose the CATALOG deliberately — that's where lock-in lives now
The 2026 decision tree is short — and its last box matters more than its first.

Day-to-day, they feel the same

The SQL is nearly interchangeable — which is itself the argument that the format war is over:

-- Iceberg (Spark SQL)
CREATE TABLE lake.sales.orders (
  order_id BIGINT, order_ts TIMESTAMP, amount DECIMAL(12,2)
) USING iceberg PARTITIONED BY (days(order_ts));

-- Time-travel predicates take literals, not computed expressions:
SELECT * FROM lake.sales.orders
FOR TIMESTAMP AS OF '2026-07-05 00:00:00';

-- Delta (Spark SQL)
CREATE TABLE lake.sales.orders (
  order_id BIGINT, order_ts TIMESTAMP, amount DECIMAL(12,2)
) USING delta;

SELECT * FROM lake.sales.orders TIMESTAMP AS OF '2026-07-05 00:00:00';

Same guarantees, slightly different spellings. Your engineers will not care which one they’re on within a month.

What they will care about is a setting both formats share and neither makes prominent: whether updates rewrite whole files or annotate them. That’s merge-on-read vs copy-on-write, and it will affect your compute bill more than the format choice on this page.

The decision that still matters: the catalog

Because the formats converged, the strategic weight moved up a layer — to the catalog that holds each table’s atomic pointer and governs who may commit: Unity (Databricks), Apache Polaris and Snowflake Open Catalog (the same codebase, self-hosted and managed), Glue (AWS), Nessie, and others. Snowflake Horizon, often listed alongside these, is the governance suite rather than the catalog itself — a distinction worth keeping straight when you compare offerings. The catalog is where access control, auditing, and cross-engine governance actually happen, and switching catalogs is far more painful than translating table metadata. If you’re building a lakehouse in 2026, and the two head-to-heads that come up most are Unity Catalog vs Glue and Polaris vs Nessie. Spend your evaluation weeks there — choosing the catalog is its own decision, and the format choice, pleasantly, has become the easy part.

Common questions

Which is better, Apache Iceberg or Delta Lake?

Neither, on raw capabilities — both provide ACID transactions, schema evolution, time travel, and deletion vectors. Delta introduced deletion vectors incrementally from 2.3.0 in 2023, with full MERGE support and on-by-default behaviour landing in 3.1.0; Iceberg standardised them in the v3 spec. The practical answer is platform fit: Delta is the frictionless choice inside Databricks and Fabric; Iceberg is the safer default for open, multi-engine lakehouses because every major platform reads and writes it.

What is Delta UniForm?

A Delta Lake feature that writes Iceberg (and Hudi) metadata alongside Delta's own, so one physical table can be read by engines expecting any of the three formats. It's Databricks' answer to Iceberg's rise as the neutral standard — write Delta, present Iceberg.

Is the Iceberg vs Delta decision permanent?

Much less than it used to be. UniForm and Apache XTable can translate metadata between formats over the same Parquet files, so a migration no longer means rewriting data. Optimize for your dominant platform today and let interoperability cover the rest.

What matters more than the table format?

The catalog — the component that holds each table's pointer and decides who may commit to it. Formats have converged; catalogs (Unity, Polaris, Snowflake Open Catalog, Glue, Nessie) are where governance and the remaining lock-in now live. Choose the catalog as carefully as you once chose the format.