Data Lakehouse Concepts Cheat Sheet

Understand lakehouse architecture combining data lake storage with warehouse guarantees via Delta Lake, Iceberg, and Hudi table formats.

2 PagesIntermediateFeb 5, 2026

Write and Version Data with Delta Lake

Write a DataFrame as a Delta table and inspect its transaction history.

python

from delta import DeltaTabledf.write.format("delta").mode("overwrite").save("/lake/orders")# time travel to a previous versionhistory_df = spark.read.format("delta").option("versionAsOf", 3).load("/lake/orders")dt = DeltaTable.forPath(spark, "/lake/orders")dt.history().show(truncate=False)

Upsert with MERGE INTO

Apply CDC-style updates and inserts atomically using Delta's MERGE operation.

sql

MERGE INTO lake.orders AS targetUSING staging.orders_cdc AS sourceON target.order_id = source.order_idWHEN MATCHED THEN  UPDATE SET target.status = source.status, target.updated_at = source.updated_atWHEN NOT MATCHED THEN  INSERT (order_id, status, updated_at) VALUES (source.order_id, source.status, source.updated_at);

Create an Iceberg Table

Define a partitioned Iceberg table queryable from multiple engines (Spark, Trino, Flink).

sql

CREATE TABLE catalog.sales.transactions (  txn_id BIGINT,  amount DECIMAL(10,2),  event_ts TIMESTAMP)USING icebergPARTITIONED BY (days(event_ts));-- schema evolution without rewriting dataALTER TABLE catalog.sales.transactions ADD COLUMN currency STRING;

Open Table Format Comparison

Key differentiators between the three dominant lakehouse table formats.

Delta Lake- deepest Spark/Databricks integration, ACID via transaction log
Apache Iceberg- strongest multi-engine support (Trino, Flink, Spark) and hidden partitioning
Apache Hudi- optimized for frequent upserts/incremental ingestion pipelines
ACID transactions- atomic commits so readers never see partial writes
Time travel- query a table as of a prior version or timestamp
Schema evolution- add/rename/drop columns without rewriting historical files

Pro Tip

Pick a table format based on which query engines your organization actually runs — Iceberg's multi-engine portability matters far more than benchmark differences once you have Trino, Spark, and Flink all reading the same tables.

Was this cheat sheet helpful?

AutoML Basics Cheat Sheet

Feature Stores Cheat Sheet