Data platform comparison / architecture decisions
Apache Iceberg vs Delta Lake: Architecture, Compatibility, UniForm and Fabric
The difficult question is not whether Iceberg or Delta has ACID transactions. It is whether every engine in your design can correctly read, write, secure and maintain the exact table features you enable.
1. Architecture: table state is more than data files
| Dimension | Apache Iceberg | Delta Lake |
|---|---|---|
| Table state | Metadata tracks schemas, partition specs and snapshots. In adopted v1–v3 layouts, snapshots reference manifest lists and manifests that enumerate table files. | The transaction log under _delta_log tracks protocol, metadata and file actions; checkpoints accelerate reconstruction of a table snapshot. |
| Data formats | The specification supports Parquet, Avro and ORC, subject to engine/feature support. | Data files are Parquet, with required log metadata and supported table features. |
| Commit coordination | Atomic metadata update/commit semantics implemented by the supported catalog or table implementation. | Transaction-log commit semantics implemented by compatible writers and supported storage/coordination. |
| Reader correctness | Must honor snapshot state, field IDs, delete representation and the table’s supported format version. | Must honor active-file actions, protocol/features, column mapping and deletion vectors when enabled. |
A folder of Parquet is not equivalent to either transactional table. Reading all files directly can include removed data, ignore deletes, duplicate rewritten rows or misinterpret physical column names. A ‘compatible engine’ must support the table protocol and features—not merely open Parquet.
A catalog is important for discovery and governed access, but not every catalog is the same API or security authority. Delta path-based reads do not require a separate catalog registration in all deployments; they still require supported storage access and a correct Delta reader.
2. Compare features at the format-and-engine level
| Concern | What to check |
|---|---|
| Schema evolution | Iceberg uses field IDs and supports documented metadata evolution. Delta supports evolution and column mapping for operations such as rename/drop. Check reader, writer and streaming compatibility before enabling features. |
| Partition/layout changes | Iceberg supports partition evolution and hidden partitioning. Delta layout capabilities depend on the implementation/features; do not substitute a vendor-specific clustering feature for a universal format capability. |
| Updates and deletes | Determine whether the engine writes rewritten files, position/equality deletes or supported deletion vectors. Each reader must apply the actual representation. |
| Time travel | History needs both metadata and referenced files. Snapshot expiration or VACUUM can make older states unavailable. |
| Streaming and CDC | Check the connector, operation semantics, checkpointing, deletes, schema changes and recovery—not a blanket format ranking. |
| Transactions | Per-table ACID does not automatically provide atomic publication across several tables or engines. |
Both formats support transactional analytics, but engine integrations can have different read and write coverage. An engine that can SELECT from a table might not support MERGE, schema evolution or concurrent commits to that table.
Column mapping is a good example of why broad rankings mislead: Delta can support metadata-based rename/drop, but enabling that feature changes compatibility requirements. ‘Delta cannot rename columns safely’ is not a useful general rule.
3. Separate the Iceberg specification from a vendor release
The Apache Iceberg specification currently identifies versions 1, 2 and 3 as complete/adopted, and version 4 as under active development rather than formally adopted. v2 introduced row-level delete files; v3 extends capabilities including deletion vectors, row lineage, default values and additional types.
That does not mean every Spark, Flink, Trino, Snowflake, BigQuery, Databricks or Fabric path implements every v3 feature. Check the product/runtime/connector matrix for the specific operation. A preview announcement in one vendor is not the global adoption status of the Iceberg specification.
Likewise, convergence proposals or a future Delta release name do not establish current interoperability. Row lineage metadata is not the same promise as end-to-end business provenance through every transformation. Do not enable a newer table version simply because one writer supports it if a required reader does not.
4. Multi-engine writing needs a supported commit path
Iceberg’s atomic metadata swap and optimistic concurrency model allow supported writers to detect changes and retry compatible operations. This is not a guarantee that every listed engine can write through any REST catalog with identical isolation and feature support.
Delta also supports concurrent writers on supported storage through optimistic concurrency control. Conflicting updates, schema/protocol changes and overlapping maintenance can fail and require a correct retry or redesign. It is not accurate to reduce Delta to ‘single-engine writes only’.
- Identify one authoritative table/catalog and supported commit protocol.
- List every writer, its version, credentials and operations.
- Verify conflicts, retries, delete semantics and schema changes using the actual clients.
- Ensure maintenance does not delete files still needed by readers or pending writes.
- Never let one writer update only Iceberg metadata while another independently updates only a Delta log over shared files unless that exact coordinated architecture is supported.
REST catalog compatibility alone does not prove shared authorization, credential vending, row-filter enforcement, write support or atomic multi-table transactions. Those require separate evidence.
5. Delta UniForm: useful read interoperability, with limits
Azure Databricks documents Iceberg reads on Delta Lake (UniForm) as generating Iceberg metadata asynchronously alongside the Delta log. Eligible Iceberg clients can read the same underlying data files. They are read-only clients through this feature; independent Iceberg writes are not supported.
The documented IcebergCompatV2 path requires Unity Catalog registration, column mapping, the specified Delta reader/writer protocol and supported Databricks Runtime writes. The current guide states deletion vectors are not supported with Iceberg reads; use its documented REORG process when conversion requires rewriting or purging incompatible features.
Do not assume that Iceberg v3’s existence automatically removes a limitation stated by the UniForm implementation you are using. Follow the actual supported feature path rather than combining two separate announcements.
Metadata generation is asynchronous. Check conversion status and the latest converted Delta version before using the external reader for freshness-sensitive work. Iceberg snapshot IDs and Delta versions are not guaranteed to match. Some enable/upgrade operations can rewrite data; ‘no copy’ does not mean every conversion is free of compute or file rewrites.
Client-specific limits still apply. A table being readable by one Iceberg engine does not certify every engine, data type or security model.
6. Fabric and OneLake: distinguish native writes from virtual reads
OneLake supports metadata virtualization between eligible Delta and Iceberg tables. A supported Iceberg table or shortcut can be presented with virtual Delta metadata; a Delta table can expose virtual Iceberg metadata. Inspect conversion logs, supported types, delete features and freshness rather than promising immediate universal compatibility.
Microsoft also documents native Iceberg table creation/commit operations through the OneLake Iceberg metadata API. This is a different write path from virtual Iceberg metadata generated for a Delta table.
| OneLake path | Supported meaning |
|---|---|
| Native Iceberg table/catalog commit | A supported client writes data/metadata and commits through the documented API with concurrency checks. |
| Delta table exposed as virtual Iceberg | Read compatibility; the virtual representation does not accept Iceberg catalog commits. Update the source using a supported Delta writer. |
| Table shortcut | Access/reference and virtualization as supported; not permission to commit arbitrary Iceberg table updates through it. |
Current OneLake Iceberg guidance lists translation limitations, including v3 metadata/type features that do not all translate. OneLake security also depends on the client and authorized enforcement path; a catalog endpoint alone does not prove every external engine enforces every row/column policy.
Fabric Lakehouse Delta tables remain a normal choice for Fabric Spark and supported Direct Lake use. That does not justify saying every Fabric workload stores its data identically or recommending Databricks UniForm as the configuration procedure for Fabric tables.
7. Choose and test before migrating
| Your situation | Practical approach |
|---|---|
| Existing reliable Fabric/Databricks Delta pipeline | Retain it unless a requirement is unmet. Test a supported read bridge before a format migration. |
| Several engines with a required Iceberg write ecosystem | Evaluate the exact Iceberg version, catalog and each writer/reader—not an abstract vendor list. |
| Only one additional reader is needed | Test supported federation, metadata virtualization, sharing or materialization; a full rewrite may be unnecessary. |
| Regulated deletes or sensitive data | Test delete visibility, retention, identity and every direct-access path before choosing an interoperability layer. |
A representative acceptance test should include insert/update/delete, exact aggregates, decimal/timestamp/nested types, schema rename/drop, restart/retry and concurrent modification. Record each reader’s observed snapshot/version. Include a denied-access test; a correct total is not a security test.
Test retention and recovery separately. Delta VACUUM and Iceberg snapshot expiration/orphan cleanup are not harmless housekeeping commands. A time-travel promise is valid only while required files and metadata remain. Keep recoverability and lagging readers within the maintenance policy.
If migration is still required, inventory unsupported features first, choose a documented conversion or rewrite path, reconcile data and consumer behavior, then plan cutover and rollback. Metadata translation is not automatically a transfer of catalog policies, streaming checkpoints, history or ownership. Neither format ‘wins’ every workload.
Common questions
Are Iceberg and Delta identical because both can use Parquet?
No. Table metadata, transaction state, delete semantics, types and protocol features determine correctness. Iceberg can also use other specified data-file formats where the engine supports them.
Can every engine write an Iceberg table through any REST catalog?
No. Check exact engine/catalog versions, operations, storage credentials, commit semantics and supported table features. Read support is not write support.
Does UniForm allow Iceberg clients to write Delta tables?
Not through the documented Iceberg-reads feature. It provides asynchronous Iceberg read metadata for eligible Delta tables; external Iceberg writes are not supported by that path.
Is Iceberg v3 support universal?
No. The adopted specification and each vendor/runtime/connector implementation have different scope. Check every required reader, writer and maintenance tool.
Can OneLake accept Iceberg writes to a virtual Delta representation?
No. The documented native Iceberg commit path is distinct. Virtual Iceberg metadata for Delta provides read compatibility; update the Delta source through a compatible Delta writer.
Official references and review scope
Checked against the following official vendor documentation on 2026-10-30. This is an independent comparison, not a vendor endorsement or a benchmark report. Recommendations are conditional on workload requirements. No customer deployment, savings result or performance benchmark is claimed; validate the design and current feature availability in your environment.
- Apache Iceberg specification and version status
- Iceberg schema and partition evolution
- Iceberg snapshot expiration and maintenance
- Delta Lake table reads/writes
- Delta concurrency control
- Delta column mapping and reader implications
- Azure Databricks UniForm requirements and limitations
- OneLake Iceberg virtualization and limitations
- OneLake Iceberg native versus virtual commit support



