Skip to main content
Version: Current

Spark Quick Start

This guide provides a quick peek at Hudi's capabilities using Spark. Using Spark Datasource APIs(both scala and python) and using Spark SQL, we will walk through code snippets that allows you to insert, update, delete and query a Hudi table.

Setup​

Hudi works with Spark 3.3 and above versions. You can follow instructions here for setting up Spark.

Spark Support Matrix​

HudiSupported Spark versionsScala versionsJava versions
1.1.x4.0.x, 3.5.x (default build), 3.4.x, 3.3.x2.13 (Spark 4.0), 2.12/2.13 (Spark 3.5), 2.12 (Spark 3.3-3.4)17+ (Spark 4.0), 8+ (Spark 3.x)
1.0.x3.5.x (default build), 3.4.x, 3.3.x2.12/2.13 (Spark 3.5), 2.12 (Spark 3.3-3.4)8+
0.15.x3.5.x (default build), 3.4.x, 3.3.x, 3.2.x, 3.1.x, 3.0.x2.12/2.13 (Spark 3.5), 2.128+
0.14.x3.4.x (default build), 3.3.x, 3.2.x, 3.1.x, 3.0.x2.128+
note

The default build Spark version indicates how we build hudi-spark3-bundle.

Spark Shell/SQL​

From the extracted directory run spark-shell with Hudi:

# For Spark versions: 3.3 - 4.0
export SPARK_VERSION=3.5
export HUDI_VERSION=1.1.0
# For Scala versions: 2.12/2.13
export SCALA_VERSION=2.13

spark-shell --master "local[2]" \
  --packages org.apache.hudi:hudi-spark$SPARK_VERSION-bundle_$SCALA_VERSION:$HUDI_VERSION \
  --conf 'spark.serializer=org.apache.spark.serializer.KryoSerializer' \
  --conf 'spark.sql.catalog.spark_catalog=org.apache.spark.sql.hudi.catalog.HoodieCatalog' \
  --conf 'spark.sql.extensions=org.apache.spark.sql.hudi.HoodieSparkSessionExtension' \
  --conf 'spark.kryo.registrator=org.apache.spark.HoodieSparkKryoRegistrar'

Note: You must adjust the SPARK_VERSION and SCALA_VERSION variables based on your environment requirements.

on Kryo serialization

Users are recommended to set this config to reduce Kryo serialization overhead

--conf 'spark.kryo.registrator=org.apache.spark.HoodieSparkKryoRegistrar'

Setup project​

Below, we do imports and setup the table name and corresponding base path.

// spark-shell

val tableName = "trips_table"
val basePath = "file:///tmp/trips_table"

Create Table​

First, let's create a Hudi table. Here, we use a partitioned table for illustration, but Hudi also supports non-partitioned tables.

// scala
// First commit will auto-initialize the table, if it did not exist in the specified base path.

Insert data​

Generate some new records as a DataFrame and write the DataFrame into a Hudi table. Since, this is the first write, it will also auto-create the table.

// spark-shell
val columns = Seq("ts","uuid","rider","driver","fare","city")
val data =
  Seq((1695159649087L,"334e26e9-8355-45cc-97c6-c31daf0df330","rider-A","driver-K",19.10,"san_francisco"),
    (1695091554788L,"e96c4396-3fad-413a-a942-4cb36106d721","rider-C","driver-M",27.70 ,"san_francisco"),
    (1695046462179L,"9909a8b1-2d15-4d3d-8ec9-efc48c536a00","rider-D","driver-L",33.90 ,"san_francisco"),
    (1695516137016L,"e3cf430c-889d-4015-bc98-59bdce1e530c","rider-F","driver-P",34.15,"sao_paulo"    ),
    (1695115999911L,"c8abbe79-8d89-47ea-b4ce-4d224bae5bfa","rider-J","driver-T",17.85,"chennai"));

var inserts = spark.createDataFrame(data).toDF(columns:_*)
inserts.write.format("hudi").
  option("hoodie.datasource.write.partitionpath.field", "city").
  option("hoodie.table.name", tableName).
  mode("overwrite").
  save(basePath)
Mapping to Hudi write operations

Hudi provides a wide range of write operations - both batch and incremental - to write data into Hudi tables, with different semantics and performance. When record keys are not configured (see keys below), bulk_insert will be chosen as the write operation, matching the out-of-behavior of Spark's Parquet Datasource.

Query data​

Hudi tables can be queried back into a DataFrame or Spark SQL.

// spark-shell
val tripsDF = spark.read.format("hudi").load(basePath)
tripsDF.createOrReplaceTempView("trips_table")

spark.sql("SELECT uuid, fare, ts, rider, driver, city FROM  trips_table WHERE fare > 20.0").show()
spark.sql("SELECT _hoodie_commit_time, _hoodie_record_key, _hoodie_partition_path, rider, driver, fare FROM  trips_table").show()

Update data​

Hudi tables can be updated by streaming in a DataFrame or using a standard UPDATE statement.

// Lets read data from target Hudi table, modify fare column for rider-D and update it. 
val updatesDf = spark.read.format("hudi").load(basePath).filter($"rider" === "rider-D").withColumn("fare", col("fare") * 10)

updatesDf.write.format("hudi").
  option("hoodie.datasource.write.operation", "upsert").
  option("hoodie.datasource.write.partitionpath.field", "city").
  option("hoodie.table.name", tableName).
  mode("append").
  save(basePath)
Key requirements

Updates with spark-datasource is feasible only when the source dataframe contains Hudi's meta fields or a key field is configured. Notice that the save mode is now Append. In general, always use append mode unless you are trying to create the table for the first time.

Querying the data again will now show updated records. Each write operation generates a new commit. Look for changes in _hoodie_commit_time, fare fields for the given _hoodie_record_key value from a previous commit.

Merging Data​

// spark-shell
Feel free to use "upsert" operation as showed under "Update data" section. Or leverage MergeInto with Spark sql writes.

Delete data​

Delete operation removes the records specified from the table. For example, this code snippet deletes records for the HoodieKeys passed in. Check out the deletion section for more details.

// spark-shell
// Lets  delete rider: rider-D
val deletesDF = spark.read.format("hudi").load(basePath).filter($"rider" === "rider-F")

deletesDF.write.format("hudi").
  option("hoodie.datasource.write.operation", "delete").
  option("hoodie.datasource.write.partitionpath.field", "city").
  option("hoodie.table.name", tableName).
  mode("append").
  save(basePath)

Querying the data again will not show the deleted record.

Key requirements

Deletes with spark-datasource is supported only when the source dataframe contains Hudi's meta fields or a key field is configured. Notice that the save mode is again Append.

Index data​

Hudi supports indexing on columns to speed up queries. Indexes can be created on columns using the CREATE INDEX statement.

note

Please note in order to create secondary index:

  1. The table must have a primary key and merge mode should be COMMIT_TIME_ORDERING.
  2. Record index must be enabled. This can be done by setting hoodie.metadata.record.index.enable=true and then creating record_index. Please note the example below.
-- Create a table with primary key
CREATE TABLE hudi_indexed_table (
    ts BIGINT,
    uuid STRING,
    rider STRING,
    driver STRING,
    fare DOUBLE,
    city STRING
) USING HUDI
options(
    primaryKey ='uuid',
    hoodie.write.record.merge.mode = "COMMIT_TIME_ORDERING"
)
PARTITIONED BY (city);

INSERT INTO hudi_indexed_table
VALUES
(1695159649,'334e26e9-8355-45cc-97c6-c31daf0df330','rider-A','driver-K',19.10,'san_francisco'),
(1695091554,'e96c4396-3fad-413a-a942-4cb36106d721','rider-C','driver-M',27.70 ,'san_francisco'),
(1695046462,'9909a8b1-2d15-4d3d-8ec9-efc48c536a00','rider-D','driver-L',33.90 ,'san_francisco'),
(1695332066,'1dced545-862b-4ceb-8b43-d2a568f6616b','rider-E','driver-O',93.50,'san_francisco'),
(1695516137,'e3cf430c-889d-4015-bc98-59bdce1e530c','rider-F','driver-P',34.15,'sao_paulo'    ),
(1695376420,'7a84095f-737f-40bc-b62f-6b69664712d2','rider-G','driver-Q',43.40 ,'sao_paulo'    ),
(1695173887,'3eeb61f7-c2b0-4636-99bd-5d7a5a1d2c04','rider-I','driver-S',41.06 ,'chennai'      ),
(1695115999,'c8abbe79-8d89-47ea-b4ce-4d224bae5bfa','rider-J','driver-T',17.85,'chennai');

-- Setting the Lock Provider
SET hoodie.write.lock.provider = org.apache.hudi.client.transaction.lock.InProcessLockProvider;
-- Create bloom filter expression index on driver column
CREATE INDEX idx_bloom_driver ON hudi_indexed_table USING bloom_filters(driver) OPTIONS(expr='identity');
-- It would show bloom filter expression index
SHOW INDEXES FROM hudi_indexed_table;
-- Query on driver column would prune the data using the idx_bloom_driver index
SELECT uuid, rider FROM hudi_indexed_table WHERE driver = 'driver-S';

-- Create column stat expression index on ts column
CREATE INDEX idx_column_ts ON hudi_indexed_table USING column_stats(ts) OPTIONS(expr='from_unixtime', format = 'yyyy-MM-dd');
-- Shows both expression indexes
SHOW INDEXES FROM hudi_indexed_table;
-- Query on ts column would prune the data using the idx_column_ts index
SELECT * FROM hudi_indexed_table WHERE from_unixtime(ts, 'yyyy-MM-dd') = '2023-09-24';

-- To create secondary index, first create the record index
SET hoodie.metadata.record.index.enable=true;
CREATE INDEX record_index ON hudi_indexed_table (uuid);
-- Create secondary index on rider column
CREATE INDEX idx_rider ON hudi_indexed_table (rider);

-- Expression index and secondary index should show up
SHOW INDEXES FROM hudi_indexed_table;
-- Query on rider column would leverage the secondary index idx_rider
SELECT * FROM hudi_indexed_table WHERE rider = 'rider-E';

-- Update a record and query the table based on indexed columns
UPDATE hudi_indexed_table SET rider = 'rider-B', driver = 'driver-N', ts = '1697516137' WHERE rider = 'rider-A';
-- Data skipping would be performed using column stat expression index
SELECT uuid, rider FROM hudi_indexed_table WHERE from_unixtime(ts, 'yyyy-MM-dd') = '2023-10-17';
-- Data skipping would be performed using bloom filter expression index
SELECT * FROM hudi_indexed_table WHERE driver = 'driver-N';
-- Data skipping would be performed using secondary index
SELECT * FROM hudi_indexed_table WHERE rider = 'rider-B';

-- Drop all the indexes
DROP INDEX record_index on hudi_indexed_table;
DROP INDEX secondary_index_idx_rider on hudi_indexed_table;
DROP INDEX expr_index_idx_bloom_driver on hudi_indexed_table;
DROP INDEX expr_index_idx_column_ts on hudi_indexed_table;
-- No indexes should show up for the table
SHOW INDEXES FROM hudi_indexed_table;

SET hoodie.metadata.record.index.enable=false;

Time Travel Query​

Hudi supports time travel query to query the table as of a point-in-time in history. Three timestamp formats are supported as illustrated below.

spark.read.format("hudi").
  option("as.of.instant", "20210728141108100").
  load(basePath)

spark.read.format("hudi").
  option("as.of.instant", "2021-07-28 14:11:08.200").
  load(basePath)

// It is equal to "as.of.instant = 2021-07-28 00:00:00"
spark.read.format("hudi").
  option("as.of.instant", "2021-07-28").
  load(basePath)

Incremental query​

Hudi provides the unique capability to obtain a set of records that changed between a start and end commit time, providing you with the "latest state" for each such record as of the end commit time. By default, Hudi tables are configured to support incremental queries, using record level metadata tracking.

Below, we fetch changes since a given begin time while the end time defaults to the latest commit on the table. Users can also specify an end time using END_INSTANTTIME.key() option.

// spark-shell
spark.read.format("hudi").load(basePath).createOrReplaceTempView("trips_table")

val commits = spark.sql("SELECT DISTINCT(_hoodie_commit_time) AS commitTime FROM  trips_table ORDER BY commitTime").map(k => k.getString(0)).take(50)
val beginTime = commits(commits.length - 2) // commit time we are interested in

// incrementally query data
val tripsIncrementalDF = spark.read.format("hudi").
  option("hoodie.datasource.query.type", "incremental").
  option("hoodie.datasource.read.begin.instanttime", 0).
  load(basePath)
tripsIncrementalDF.createOrReplaceTempView("trips_incremental")

spark.sql("SELECT `_hoodie_commit_time`, fare, rider, driver, uuid, ts FROM  trips_incremental WHERE fare > 20.0").show()

Change Data Capture Query​

Hudi also exposes first-class support for Change Data Capture (CDC) queries. CDC queries are useful for applications that need to obtain all the changes, along with before/after images of records, given a commit time range.

// spark-shell
// Lets first insert data to a new table with cdc enabled.
val columns = Seq("ts","uuid","rider","driver","fare","city")
val data =
  Seq((1695158649187L,"334e26e9-8355-45cc-97c6-c31daf0df330","rider-A","driver-K",19.10,"san_francisco"),
    (1695091544288L,"e96c4396-3fad-413a-a942-4cb36106d721","rider-B","driver-L",27.70 ,"san_paulo"),
    (1695046452379L,"9909a8b1-2d15-4d3d-8ec9-efc48c536a00","rider-C","driver-M",33.90 ,"san_francisco"),
    (1695332056404L,"1dced545-862b-4ceb-8b43-d2a568f6616b","rider-D","driver-N",93.50,"chennai"));
var df = spark.createDataFrame(data).toDF(columns:_*)

// Insert data
df.write.format("hudi").
  option("hoodie.datasource.write.partitionpath.field", "city").
  option("hoodie.table.cdc.enabled", "true").
  option("hoodie.table.name", tableName).
  mode("overwrite").
  save(basePath)

// Update fare for riders: rider-A and rider-B 
val updatesDf = spark.read.format("hudi").load(basePath).filter($"rider" === "rider-A" || $"rider" === "rider-B").withColumn("fare", col("fare") * 10)

updatesDf.write.format("hudi").
  option("hoodie.datasource.write.operation", "upsert").
  option("hoodie.datasource.write.partitionpath.field", "city").
  option("hoodie.table.cdc.enabled", "true").
  option("hoodie.table.name", tableName).
  mode("append").
  save(basePath)


// Query CDC data
spark.read.option("hoodie.datasource.read.begin.instanttime", 0).
  option("hoodie.datasource.query.type", "incremental").
  option("hoodie.datasource.query.incremental.format", "cdc").
  format("hudi").load(basePath).show(false)
Key requirements

Note that CDC queries are currently only supported on Copy-on-Write tables.

Table Types​

The examples thus far have showcased one of the two table types, that Hudi supports - Copy-on-Write (COW) tables. Hudi also supports a more advanced write-optimized table type called Merge-on-Read (MOR) tables, that can balance read and write performance in a more flexible manner. See table types for more details.

Any of these examples can be run on a Merge-on-Read table by simply changing the table type to MOR, while creating the table, as below.

// spark-shell
inserts.write.format("hudi").
  ...
  option("hoodie.datasource.write.table.type", "MERGE_ON_READ").
  ...

Keys​

Hudi also allows users to specify a record key, which will be used to uniquely identify a record within a Hudi table. This is useful and critical to support features like indexing and clustering, which speed up upserts and queries respectively, in a consistent manner. Some of the other benefits of keys are explained in detail here. To this end, Hudi supports a wide range of built-in key generators, that make it easy to generate record keys for a given table. In the absence of a user configured key, Hudi will auto generate record keys, which are highly compressible.

// spark-shell
inserts.write.format("hudi").
...
option("hoodie.datasource.write.recordkey.field", "uuid").
...
Implications of defining record keys

Configuring keys for a Hudi table, has a new implications on the table. If record key is set by the user, upsert is chosen as the write operation. Also if a record key is configured, then it's also advisable to specify ordering fields, to correctly handle cases where the source data has multiple records with the same key. See section below.

Merge Modes​

Hudi also allows users to specify ordering fields, which will be used to order and resolve conflicts between multiple versions of the same record. This is very important for use-cases like applying database CDC logs to a Hudi table, where a given record may appear multiple times in the source data due to repeated upstream updates. Hudi also uses this mechanism to support out-of-order data arrival into a table, where records may need to be resolved in a different order than their commit time. For e.g. using a created_at timestamp field as an ordering field will prevent older versions of a record from overwriting newer ones or being exposed to queries, even if they are written at a later commit time to the table. This is one of the key features, that makes Hudi, best suited for dealing with streaming data.

To enable different merge semantics, Hudi supports merge modes. Commit time and event time based merge modes are supported out of the box. Users can also define their own custom merge strategies, see here.

// spark-shell 
updatesDf.write.format("hudi").
  ...
  option("hoodie.table.ordering.fields", "ts").
  ...

Where to go from here?​

You can also build hudi yourself and try this quickstart using --jars <path to spark bundle jar>(see also build with scala 2.12) for more info. If you are looking for ways to migrate your existing data to Hudi, refer to migration guide.

Spark SQL Reference​

For advanced usage of spark SQL, please refer to Spark SQL DDL and Spark SQL DML reference guides. For alter table commands, check out this. Stored procedures provide a lot of powerful capabilities using Hudi SparkSQL to assist with monitoring, managing and operating Hudi tables, please check this out.

Streaming workloads​

Hudi provides industry-leading performance and functionality for streaming data.

Hudi Streamer - Hudi provides an incremental ingestion/ETL tool - Hudi Streamer, to assist with ingesting data into Hudi from various different sources in a streaming manner, with powerful built-in capabilities like auto checkpointing, schema enforcement via schema provider, transformation support, automatic table services and so on.

Structured Streaming - Hudi supports Spark Structured Streaming reads and writes as well. Please see here for more.

Check out more information on modeling data in Hudi and different ways to perform batch writes and streaming writes.

Dockerized Demo​

Even as we showcased the core capabilities, Hudi supports a lot more advanced functionality that can make it easy to get your transactional data lakes up and running quickly, across a variety query engines like Hive, Flink, Spark, Presto, Trino and much more. We have put together a demo video that showcases all of this on a docker based setup with all dependent systems running locally. We recommend you replicate the same setup and run the demo yourself, by following steps here to get a taste for it.