Use it when

Your platform is Spark- or Databricks-centric; Delta is the default format there and the integration is the deepest available.
You need ACID transactions, schema enforcement, upserts, and time travel over plain Parquet with no separate catalog service.
Downstream consumers read the Change Data Feed rather than rescanning tables.
Update-heavy tables benefit from deletion vectors, which avoid rewriting whole files on merge.

Think twice when

Writes come mostly from engines other than Spark; Trino, Flink, and the standalone readers are real but less complete than the Spark path.
Multiple clusters write to one table on S3, which needs a DynamoDB-backed LogStore or coordinated commits.
Renaming or dropping columns is routine; both require column mapping to be enabled first.
The rest of your stack has standardized on Iceberg; ecosystem alignment is usually the deciding factor between the two.

How it runs

Delta is a library in the engine plus a transaction log (_delta_log) stored inside the table path, so a catalog is optional. Data is Parquet on object storage; commits append to the log with optimistic concurrency. You run OPTIMIZE to compact and VACUUM to expire files; log checkpoints are automatic. Databricks operates all of this as a managed service, and UniForm can write Iceberg and Hudi metadata alongside for outside readers.

Details

Compare

How Delta Lake answers the questions Table Formats turns on.

Table Formats

How it works

Catalog

Optional: _delta_log in the table path

Update model

Copy-on-write; merge-on-read with deletion vectors

Partitioning

Hive-style columns, or liquid clustering with CLUSTER BY, up to four keys

Schema evolution

Add and reorder freely; rename and drop need column mapping enabled

Concurrent writers

Optimistic log commits; multi-cluster S3 needs a DynamoDB LogStore or coordinated commits

Indexes

File stats for data skipping, Z-order and liquid clustering; bloom filters on Databricks

Change feed

Change Data Feed, written to _change_data

Running it

Maintenance

OPTIMIZE to compact, VACUUM to expire files; log checkpoints are automatic

Connections

Engine support

Spark-first; Trino, Flink and standalone readers

Cross-format reads

UniForm writes Iceberg and Hudi metadata alongside, GA in 4.0

Alternatives in Table Formats

View all technologies

Apache Hudi

Transactional lake format built around upserts and incremental reads

vs Delta Lake: Java

Apache Hudi is an open lakehouse format that pioneered transactional data lakes, offering record-level upserts and deletes, incremental queries, and a choice between copy-on-write and merge-on-read table types.

Apache Iceberg

Open table format for large analytic datasets

vs Delta Lake: Java, Python, Rust, Go

Apache Iceberg is an open table format bringing ACID transactions, schema evolution, and time travel to data lake tables, with broad engine support.

Apache Paimon

Lake format with LSM-backed streaming updates and changelogs

vs Delta Lake: Java

Apache Paimon is a lake format unifying batch and streaming, using an LSM tree to absorb high-frequency updates and emitting changelogs so downstream jobs can consume what changed rather than rescanning.