Querying Data

Conceptually, Hudi stores data physically once on DFS, while providing 3 different ways of querying, as explained before. Once the table is synced to the Hive metastore, it provides external Hive tables backed by Hudi's custom inputformats. Once the proper hudi bundle has been installed, the table can be queried by popular query engines like Hive, Spark SQL, Spark Datasource API and PrestoDB.

In sections, below we will discuss specific setup to access different query types from different query engines.

Spark Datasource

The Spark Datasource API is a popular way of authoring Spark ETL pipelines. Hudi tables can be queried via the Spark datasource with a simple spark.read.parquet. See the Spark Quick Start for more examples of Spark datasource reading queries.

To setup Spark for querying Hudi, see the Query Engine Setup page.

Snapshot query

Retrieve the data table at the present point in time.

val hudiIncQueryDF = spark
     .read()
     .format("hudi")
     .option(DataSourceReadOptions.QUERY_TYPE_OPT_KEY(), DataSourceReadOptions.QUERY_TYPE_SNAPSHOT_OPT_VAL())
     .load(tablePath) 

Incremental query

Of special interest to spark pipelines, is Hudi's ability to support incremental queries, like below. A sample incremental query, that will obtain all records written since beginInstantTime, looks like below. Thanks to Hudi's support for record level change streams, these incremental pipelines often offer 10x efficiency over batch counterparts by only processing the changed records.

The following snippet shows how to obtain all records changed after beginInstantTime and run some SQL on them.

 Dataset<Row> hudiIncQueryDF = spark.read()
     .format("org.apache.hudi")
     .option(DataSourceReadOptions.QUERY_TYPE_OPT_KEY(), DataSourceReadOptions.QUERY_TYPE_INCREMENTAL_OPT_VAL())
     .option(DataSourceReadOptions.BEGIN_INSTANTTIME_OPT_KEY(), <beginInstantTime>)
     .option(DataSourceReadOptions.INCR_PATH_GLOB_OPT_KEY(), "/year=2020/month=*/day=*") // Optional, use glob pattern if querying certain partitions
     .load(tablePath); // For incremental query, pass in the root/base path of table
     
hudiIncQueryDF.createOrReplaceTempView("hudi_trips_incremental")
spark.sql("select `_hoodie_commit_time`, fare, begin_lon, begin_lat, ts from  hudi_trips_incremental where fare > 20.0").show()

For examples, refer to Incremental Queries in the Spark quickstart. Please refer to configurations section, to view all datasource options.

Additionally, HoodieReadClient offers the following functionality using Hudi's implicit indexing.

API	Description
read(keys)	Read out the data corresponding to the keys as a DataFrame, using Hudi's own index for faster lookup
filterExists()	Filter out already existing records from the provided RDD[HoodieRecord]. Useful for de-duplication
checkExists(keys)	Check if the provided keys exist in a Hudi table

Spark SQL

Once the Hudi tables have been registered to the Hive metastore, they can be queried using the Spark-Hive integration. By default, Spark SQL will try to use its own parquet reader instead of Hive SerDe when reading from Hive metastore parquet tables. The following are important settings to consider when querying COPY_ON_WRITE or MERGE_ON_READ tables.

Copy On Write tables

For COPY_ON_WRITE tables, Spark's default parquet reader can be used to retain Sparks built-in optimizations for reading parquet files like vectorized reading on Hudi Hive tables. If using the default parquet reader, a path filter needs to be pushed into sparkContext as follows.

spark.sparkContext.hadoopConfiguration.setClass("mapreduce.input.pathFilter.class", classOf[org.apache.hudi.hadoop.HoodieROTablePathFilter], classOf[org.apache.hadoop.fs.PathFilter]);

Merge On Read tables

No special configurations are needed for querying MERGE_ON_READ tables with Hudi version 0.9.0+

If you are querying MERGE_ON_READ tables using Hudi version <= 0.8.0, you need to turn off the SparkSQL default parquet reader by setting: spark.sql.hive.convertMetastoreParquet=false.

$ spark-shell --driver-class-path /etc/hive/conf  --packages org.apache.hudi:hudi-spark-bundle_2.11:0.5.3,org.apache.spark:spark-avro_2.11:2.4.4 --conf spark.sql.hive.convertMetastoreParquet=false

scala> sqlContext.sql("select count(*) from hudi_trips_mor_rt where datestr = '2016-10-02'").show()
scala> sqlContext.sql("select count(*) from hudi_trips_mor_rt where datestr = '2016-10-02'").show()

note

Note: COPY_ON_WRITE tables can also still be read if you turn off the default parquet reader.

Flink SQL

Once the flink Hudi tables have been registered to the Flink catalog, it can be queried using the Flink SQL. It supports all query types across both Hudi table types, relying on the custom Hudi input formats again like Hive. Typically notebook users and Flink SQL CLI users leverage flink sql for querying Hudi tables. Please add hudi-flink-bundle as described in the Flink Quickstart.

By default, Flink SQL will try to use its own parquet reader instead of Hive SerDe when reading from Hive metastore parquet tables.

# HADOOP_HOME is your hadoop root directory after unpack the binary package.
export HADOOP_CLASSPATH=`$HADOOP_HOME/bin/hadoop classpath`

./bin/sql-client.sh embedded -j .../hudi-flink-bundle_2.1?-*.*.*.jar shell

-- this defines a COPY_ON_WRITE table named 't1'
CREATE TABLE t1(
  uuid VARCHAR(20), -- you can use 'PRIMARY KEY NOT ENFORCED' syntax to specify the field as record key
  name VARCHAR(10),
  age INT,
  ts TIMESTAMP(3),
  `partition` VARCHAR(20)
)
PARTITIONED BY (`partition`)
WITH (
  'connector' = 'hudi',
  'path' = 'table_base+path'
);

-- query the data
select * from t1 where `partition` = 'par1';

Flink's built-in support parquet is used for both COPY_ON_WRITE and MERGE_ON_READ tables, additionally partition prune is applied by Flink engine internally if a partition path is specified in the filter. Filters push down is not supported yet (already on the roadmap).

For MERGE_ON_READ table, in order to query hudi table as a streaming, you need to add option 'read.streaming.enabled' = 'true', when querying the table, a Flink streaming pipeline starts and never ends until the user cancel the job manually. You can specify the start commit with option read.start-commit and source monitoring interval with option read.streaming.check-interval.

Streaming Query

By default, the hoodie table is read as batch, that is to read the latest snapshot data set and returns. Turns on the streaming read mode by setting option read.streaming.enabled as true. Sets up option read.start-commit to specify the read start offset, specifies the value as earliest if you want to consume all the history data set.

Options

Option Name	Required	Default	Remarks
read.streaming.enabled	false	false	Specify true to read as streaming
read.start-commit	false	the latest commit	Start commit time in format 'yyyyMMddHHmmss', use earliest to consume from the start commit
read.streaming.skip_compaction	false	false	Whether to skip compaction instants for streaming read, generally for two purpose: 1) Avoid consuming duplications from compaction instants created for created by Hudi versions < 0.11.0 or when hoodie.compaction.preserve.commit.metadata is disabled 2) When change log mode is enabled, to only consume change for right semantics.
clean.retain_commits	false	10	The max number of commits to retain before cleaning, when change log mode is enabled, tweaks this option to adjust the change log live time. For example, the default strategy keeps 50 minutes of change logs if the checkpoint interval is set up as 5 minutes.

note

When option read.streaming.skip_compaction is enabled and the streaming reader lags behind by commits of number clean.retain_commits, data loss may occur.

The compaction table service action preserves the original commit time for each row. When iterating through the parquet files, the streaming reader will perform a check on whether the row's commit time falls within the specified instant range to skip over rows that have been read before.

For efficiency, option read.streaming.skip_compaction can be enabled to skip reading of parquet files entirely.

note

read.streaming.skip_compaction should only be enabled if the MOR table is compacted by Hudi with versions < 0.11.0.

This is so as the hoodie.compaction.preserve.commit.metadata feature is only introduced in Hudi versions >=0.11.0. Older versions will overwrite the original commit time for each row with the compaction plan's instant time.

This will render Hudi-on-Flink's stream reader's row-level instant-range checks to not work properly. When the original instant time is overwritten with a newer instant time, the stream reader will not be able to differentiate rows that have already been read before with actual new rows.

Incremental Query

There are 3 use cases for incremental query:

1. Streaming query: specify the start commit with option read.start-commit;
2. Batch query: specify the start commit with option read.start-commit and end commit with option read.end-commit, the interval is a closed one: both start commit and end commit are inclusive;
3. TimeTravel: consume as batch for an instant time, specify the read.end-commit is enough because the start commit is latest by default.

Options

Option Name	Required	Default	Remarks
read.start-commit	false	the latest commit	Specify earliest to consume from the start commit
read.end-commit	false	the latest commit	--

Metadata Table

The metadata table holds the metadata index per hudi table, it holds the file list and all kinds of indexes that we called multi-model index. Current these indexes are supported:

1. partition -> files mapping
2. column max/min statistics for each file
3. bloom filter for each file

The partition -> files mappings can be used for fetching the file list for writing/reading path, it is cost friendly for object storage that charges per visit, for HDFS, it can ease the access burden of the NameNode.

The column max/min statistics per file is used for query acceleration, in the writing path, when enable this feature, hudi would book-keep the max/min values for each column in real-time, thus would decrease the writing throughput. In the reading path, hudi uses this statistics to filter out the useless files first before scanning.

The bloom filter index is currently only used for spark bloom filter index, not for query acceleration yet.

In general, enable the metadata table would increase the commit time, it is not very friendly for the use cases for very short checkpoint interval (say 30s). And for these use cases you should test the stability first.

Options

Option Name	Required	Default	Remarks
metadata.enabled	false	false	Set to true to enable
read.data.skipping.enabled	false	false	Whether to enable data skipping for batch snapshot read, by default disabled
hoodie.metadata.index.column.stats.enable	false	false	Whether to enable column statistics (max/min)
hoodie.metadata.index.column.stats.column.list	false	N/A	Columns(separated by comma) to collect the column statistics

Hive

To setup Hive for querying Hudi, see the Query Engine Setup page.

Incremental query

HiveIncrementalPuller allows incrementally extracting changes from large fact/dimension tables via HiveQL, combining the benefits of Hive (reliably process complex SQL queries) and incremental primitives (speed up querying tables incrementally instead of scanning fully). The tool uses Hive JDBC to run the hive query and saves its results in a temp table. that can later be upserted. Upsert utility (HoodieDeltaStreamer) has all the state it needs from the directory structure to know what should be the commit time on the target table. e.g: /app/incremental-hql/intermediate/{source_table_name}_temp/{last_commit_included}.The Delta Hive table registered will be of the form {tmpdb}.{source_table}_{last_commit_included}.

The following are the configuration options for HiveIncrementalPuller

Config	Description	Default
hiveUrl	Hive Server 2 URL to connect to
hiveUser	Hive Server 2 Username
hivePass	Hive Server 2 Password
queue	YARN Queue name
tmp	Directory where the temporary delta data is stored in DFS. The directory structure will follow conventions. Please see the below section.
extractSQLFile	The SQL to execute on the source table to extract the data. The data extracted will be all the rows that changed since a particular point in time.
sourceTable	Source Table Name. Needed to set hive environment properties.
sourceDb	Source DB name. Needed to set hive environment properties.
targetTable	Target Table Name. Needed for the intermediate storage directory structure.
targetDb	Target table's DB name.
tmpdb	The database to which the intermediate temp delta table will be created	hoodie_temp
fromCommitTime	This is the most important parameter. This is the point in time from which the changed records are queried from.
maxCommits	Number of commits to include in the query. Setting this to -1 will include all the commits from fromCommitTime. Setting this to a value > 0, will include records that ONLY changed in the specified number of commits after fromCommitTime. This may be needed if you need to catch up say 2 commits at a time.	3
help	Utility Help

Setting fromCommitTime=0 and maxCommits=-1 will fetch the entire source table and can be used to initiate backfills. If the target table is a Hudi table, then the utility can determine if the target table has no commits or is behind more than 24 hour (this is configurable), it will automatically use the backfill configuration, since applying the last 24 hours incrementally could take more time than doing a backfill. The current limitation of the tool is the lack of support for self-joining the same table in mixed mode (snapshot and incremental modes).

NOTE on Hive incremental queries that are executed using Fetch task: Since Fetch tasks invoke InputFormat.listStatus() per partition, Hoodie metadata can be listed in every such listStatus() call. In order to avoid this, it might be useful to disable fetch tasks using the hive session property for incremental queries: set hive.fetch.task.conversion=none; This would ensure Map Reduce execution is chosen for a Hive query, which combines partitions (comma separated) and calls InputFormat.listStatus() only once with all those partitions.

PrestoDB

To setup PrestoDB for querying Hudi, see the Query Engine Setup page.

Trino

To setup Trino for querying Hudi, see the Query Engine Setup page.

Impala (3.4 or later)

Snapshot Query

Impala is able to query Hudi Copy-on-write table as an EXTERNAL TABLE on HDFS.

To create a Hudi read optimized table on Impala:

CREATE EXTERNAL TABLE database.table_name
LIKE PARQUET '/path/to/load/xxx.parquet'
STORED AS HUDIPARQUET
LOCATION '/path/to/load';

Impala is able to take advantage of the physical partition structure to improve the query performance. To create a partitioned table, the folder should follow the naming convention like year=2020/month=1. Impala use = to separate partition name and partition value.
To create a partitioned Hudi read optimized table on Impala:

CREATE EXTERNAL TABLE database.table_name
LIKE PARQUET '/path/to/load/xxx.parquet'
PARTITION BY (year int, month int, day int)
STORED AS HUDIPARQUET
LOCATION '/path/to/load';
ALTER TABLE database.table_name RECOVER PARTITIONS;

After Hudi made a new commit, refresh the Impala table to get the latest results.

REFRESH database.table_name

Redshift Spectrum

To set up Redshift Spectrum for querying Hudi, see the Query Engine Setup page.

Doris

To set up Doris for querying Hudi, see the Query Engine Setup page.

StarRocks

To set up StarRocks for querying Hudi, see the Query Engine Setup page.

Support Matrix

Following tables show whether a given query is supported on specific query engine.

Copy-On-Write tables

Query Engine	Snapshot Queries	Incremental Queries
Hive	Y	Y
Spark SQL	Y	Y
Spark Datasource	Y	Y
Flink SQL	Y	N
PrestoDB	Y	N
Trino	Y	N
Impala	Y	N
Redshift Spectrum	Y	N
Doris	Y	N
StarRocks	Y	N

Note that Read Optimized queries are not applicable for COPY_ON_WRITE tables.

Merge-On-Read tables

Query Engine	Snapshot Queries	Incremental Queries	Read Optimized Queries
Hive	Y	Y	Y
Spark SQL	Y	Y	Y
Spark Datasource	Y	Y	Y
Flink SQL	Y	Y	Y
PrestoDB	Y	N	Y
Trino	N	N	Y
Impala	N	N	Y
Redshift Spectrum	N	N	Y
StarRocks	Y	N	Y