13.2. Hive Connector

Overview

The Hive connector allows querying data stored in a Hive data warehouse. Hive is a combination of three components:

  • Data files in varying formats that are typically stored in the Hadoop Distributed File System (HDFS) or in Amazon S3.
  • Metadata about how the data files are mapped to schemas and tables. This metadata is stored in a database such as MySQL and is accessed via the Hive metastore service.
  • A query language called HiveQL. This query language is executed on a distributed computing framework such as MapReduce or Tez.

Presto only uses the first two components: the data and the metadata. It does not use HiveQL or any part of Hive’s execution environment.

Supported File Types

The following file types are supported for the Hive connector:

  • ORC
  • Parquet
  • RCFile
  • SequenceFile
  • Text

Configuration

Presto includes Hive connectors for multiple versions of Hadoop:

  • hive-hadoop2: Apache Hadoop 2.x
  • hive-cdh5: Cloudera CDH 5

Create /etc/opt/prestoadmin/connector/hive.properties with the following contents to mount the hive-hadoop2 connector as the hive catalog, replacing hive-hadoop2 with the proper connector for your version of Hadoop and example.net:9083 with the correct host and port for your Hive metastore Thrift service:

connector.name=hive-hadoop2
hive.metastore.uri=thrift://example.net:9083

Use presto-admin to deploy the connector file. See Adding a Catalog.

Multiple Hive Clusters

You can have as many catalogs as you need, so if you have additional Hive clusters, simply add another properties file to /etc/opt/prestoadmin/connector with a different name (making sure it ends in .properties). For example, if you name the property file sales.properties, Presto will create a catalog named sales using the configured connector.

HDFS Configuration

For basic setups, Presto configures the HDFS client automatically and does not require any configuration files. In some cases, such as when using federated HDFS or NameNode high availability, it is necessary to specify additional HDFS client options in order to access your HDFS cluster. To do so, add the hive.config.resources property to reference your HDFS config files:

hive.config.resources=/etc/hadoop/conf/core-site.xml,/etc/hadoop/conf/hdfs-site.xml

Only specify additional configuration files if necessary for your setup. We also recommend reducing the configuration files to have the minimum set of required properties, as additional properties may cause problems.

The configuration files must exist on all Presto nodes. If you are referencing existing Hadoop config files, make sure to copy them to any Presto nodes that are not running Hadoop.

HDFS Username

When not using Kerberos with HDFS, Presto will access HDFS using the OS user of the Presto process. For example, if Presto is running as nobody, it will access HDFS as nobody. You can override this username by setting the HADOOP_USER_NAME system property in the Presto presto_jvm_config, replacing hdfs_user with the appropriate username:

-DHADOOP_USER_NAME=hdfs_user

Accessing Hadoop clusters protected with Kerberos authentication

Kerberos authentication is currently supported for both HDFS and the Hive metastore.

However there are still a few limitations:

  • Kerberos authentication is only supported for the hive-hadoop2 and hive-cdh5 connectors.
  • Kerberos authentication by ticket cache is not yet supported.

The properties that apply to Hive connector security are listed in the Hive Configuration Properties table. Please see the Hive Security Configuration section for a more detailed discussion of the security options in the Hive connector.

HDFS Permissions

Before running any CREATE TABLE or CREATE TABLE ... AS statements for Hive tables in Presto, you need to check that the operating system user running the Presto server has access to the Hive warehouse directory on HDFS. The Hive warehouse directory is specified by the configuration variable hive.metastore.warehouse.dir in hive-site.xml, and the default value is /user/hive/warehouse. If that is not the case, either add the following to jvm.config on all of the nodes: -DHADOOP_USER_NAME=USER, where USER is an operating system user that has proper permissions for the Hive warehouse directory, or start the Presto server as a user with similar permissions. The hive user generally works as USER, since Hive is often started with the hive user. If you run into HDFS permissions problems on CREATE TABLE ... AS, remove /tmp/presto-* on HDFS, fix the user as described above, then restart all of the Presto servers.

Hive Configuration Properties

Property Name Description Default
hive.metastore.uri The URI(s) of the Hive metastore to connect to using the Thrift protocol. If multiple URIs are provided, the first URI is used by default and the rest of the URIs are fallback metastores. This property is required. Example: thrift://192.0.2.3:9083 or thrift://192.0.2.3:9083,thrift://192.0.2.4:9083  
hive.config.resources An optional comma-separated list of HDFS configuration files. These files must exist on the machines running Presto. Only specify this if absolutely necessary to access HDFS. Example: /etc/hdfs-site.xml  
hive.storage-format The default file format used when creating new tables. RCBINARY
hive.compression-codec The compression codec to use when writing files. GZIP
hive.force-local-scheduling See tuning section false
hive.respect-table-format Should new partitions be written using the existing table format or the default Presto format? true
hive.immutable-partitions Can new data be inserted into existing partitions? false
hive.max-partitions-per-writers Maximum number of partitions per writer. 100
hive.metastore.authentication.type Hive metastore authentication type. Possible values are NONE or KERBEROS. NONE
hive.metastore.service.principal The Kerberos principal of the Hive metastore service.  
hive.metastore.client.principal The Kerberos principal that Presto will use when connecting to the Hive metastore service.  
hive.metastore.client.keytab Hive metastore client keytab location.  
hive.hdfs.authentication.type HDFS authentication type. Possible values are NONE or KERBEROS. NONE
hive.hdfs.impersonation.enabled Enable HDFS end user impersonation. false
hive.hdfs.presto.principal The Kerberos principal that Presto will use when connecting to HDFS.  
hive.hdfs.presto.keytab HDFS client keytab location.  
hive.security See Hive Security Configuration.  
security.config-file Path of config file to use when hive.security=file. See File Based Authorization for details.  
hive.multi-file-bucketing.enabled Enable support for multiple files per bucket for Hive clustered tables. See Clustered Hive tables support false
hive.empty-bucketed-partitions.enabled Enable support for clustered tables with empty partitions. See Clustered Hive tables support false

Clustered Hive tables support

By default Presto supports only one data file per bucket per partition for clustered tables (Hive tables declared with CLUSTERED BY clause). If number of files does not match number of buckets exception would be thrown.

To enable support for cases where there are more than one file per bucket, when multiple INSERTs were done to a single partition of the clustered table, you can use:

  • hive.multi-file-bucketing.enabled config property
  • multi_file_bucketing_enabled session property (using SET SESSION <connector_name>.multi_file_bucketing_enabled)

Config property changes behaviour globally and session property can be used on per query basis. The default value of session property is taken from config property.

If support for multiple files per bucket is enabled Presto will group the files in partition directory. It will sort filenames lexicographically. Then it will treat part of filename up to first underscore character as bucket key. This pattern matches naming convention of files in directory when Hive is used to inject data into table.

Presto will still validate if number of file groups matches number of buckets declared for table and fail if it does not.

Similarly by default empty partitions (partitions with no files) are not allowed for clustered Hive tables. To enable support for empty paritions you can use:

  • hive.empty-bucketed-partitions.enabled config property
  • empty_bucketed_partitions_enabled session property (using SET SESSION <connector_name>.empty_bucketed_partitions_enabled)

Amazon S3 Configuration

The Hive Connector can read and write tables that are stored in S3. This is accomplished by having a table or database location that uses an S3 prefix rather than an HDFS prefix.

S3 Credentials

If you are running Presto on Amazon EC2 using EMR or another facility, it is highly recommended that you set hive.s3.use-instance-credentials to true and use IAM Roles for EC2 to govern access to S3. If this is the case, your EC2 instances will need to be assigned an IAM Role which grants appropriate access to the data stored in the S3 bucket(s) you wish to use. This is much cleaner than setting AWS access and secret keys in the hive.s3.aws-access-key and hive.s3.aws-secret-key settings, and also allows EC2 to automatically rotate credentials on a regular basis without any additional work on your part.

Custom S3 Credentials Provider

You can configure a custom S3 credentials provider by setting the Hadoop configuration property presto.s3.credentials-provider to be the fully qualified class name of a custom AWS credentials provider implementation. This class must implement the AWSCredentialsProvider interface and provide a two-argument constructor that takes a java.net.URI and a Hadoop org.apache.hadoop.conf.Configuration as arguments. A custom credentials provider can be used to provide temporary credentials from STS (using STSSessionCredentialsProvider), IAM role-based credentials (using STSAssumeRoleSessionCredentialsProvider), or credentials for a specific use case (e.g., bucket/user specific credentials). This Hadoop configuration property must be set in the Hadoop configuration files referenced by the hive.config.resources Hive connector property.

Tuning Properties

The following tuning properties affect the behavior of the client used by the Presto S3 filesystem when communicating with S3. Most of these parameters affect settings on the ClientConfiguration object associated with the AmazonS3Client.

Property Name Description Default
hive.s3.max-error-retries Maximum number of error retries, set on the S3 client. 10
hive.s3.max-client-retries Maximum number of read attempts to retry. 3
hive.s3.max-backoff-time Use exponential backoff starting at 1 second up to this maximum value when communicating with S3. 10 minutes
hive.s3.max-retry-time Maximum time to retry communicating with S3. 10 minutes
hive.s3.connect-timeout TCP connect timeout. 5 seconds
hive.s3.socket-timeout TCP socket read timeout. 5 seconds
hive.s3.max-connections Maximum number of simultaneous open connections to S3. 500
hive.s3.multipart.min-file-size Minimum file size before multi-part upload to S3 is used. 16 MB
hive.s3.multipart.min-part-size Minimum multi-part upload part size. 5 MB

S3 Data Encryption

Presto supports reading and writing encrypted data in S3 using both server-side encryption with S3 managed keys and client-side encryption using either the Amazon KMS or a software plugin to manage AES encryption keys.

With S3 server-side encryption, (called SSE-S3 in the Amazon documentation) the S3 infrastructure takes care of all encryption and decryption work (with the exception of SSL to the client, assuming you have hive.s3.ssl.enabled set to true). S3 also manages all the encryption keys for you. To enable this, set hive.s3.sse.enabled to true.

With S3 client-side encryption, S3 stores encrypted data and the encryption keys are managed outside of the S3 infrastructure. Data is encrypted and decrypted by Presto instead of in the S3 infrastructure. In this case, encryption keys can be managed either by using the AWS KMS or your own key management system. To use the AWS KMS for key management, set hive.s3.kms-key-id to the UUID of a KMS key. Your AWS credentials or EC2 IAM role will need to be granted permission to use the given key as well.

To use a custom encryption key management system, set hive.s3.encryption-materials-provider to the fully qualified name of a class which implements the EncryptionMaterialsProvider interface from the AWS Java SDK. This class will have to be accessible to the Hive Connector through the classpath and must be able to communicate with your custom key management system. If this class also implements the org.apache.hadoop.conf.Configurable interface from the Hadoop Java API, then the Hadoop configuration will be passed in after the object instance is created and before it is asked to provision or retrieve any encryption keys.

Schema Evolution

Hive allows the partitions in a table to have a different schema than the table. This occurs when the column types of a table are changed after partitions already exist (that use the original column types). The Hive connector supports this by allowing the same conversions as Hive:

  • varchar to and from tinyint, smallint, integer and bigint
  • real to double
  • Widening conversions for integers, such as tinyint to smallint

Any conversion failure will result in null, which is the same behavior as Hive. For example, converting the string 'foo' to a number, or converting the string '1234' to a tinyint (which has a maximum value of 127).

Examples

The Hive connector supports querying and manipulating Hive tables and schemas (databases). While some uncommon operations will need to be performed using Hive directly, most operations can be performed using Presto.

Create a new Hive schema named web that will store tables in an S3 bucket named my-bucket:

CREATE SCHEMA hive.web
WITH (location = 's3://my-bucket/')

Create a new Hive table named page_views in the web schema that is stored using the ORC file format, partitioned by date and country, and bucketed by user into 50 buckets (note that Hive requires the partition columns to be the last columns in the table):

CREATE TABLE hive.web.page_views (
  view_time timestamp,
  user_id bigint,
  page_url varchar,
  ds date,
  country varchar
)
WITH (
  format = 'ORC',
  partitioned_by = ARRAY['ds', 'country'],
  bucketed_by = ARRAY['user_id'],
  bucket_count = 50
)

Drop a partition from the page_views table:

DELETE FROM hive.web.page_views
WHERE ds = DATE '2016-08-09'
  AND country = 'US'

Query the page_views table:

SELECT * FROM hive.web.page_views

Create an external Hive table named request_logs that points at existing data in S3:

CREATE TABLE hive.web.request_logs (
  request_time timestamp,
  url varchar,
  ip varchar,
  user_agent varchar
)
WITH (
  format = 'TEXTFILE',
  external_location = 's3://my-bucket/data/logs/'
)

Drop the external table request_logs. This only drops the metadata for the table. The referenced data directory is not deleted:

DROP hive.web.request_logs

Drop a schema:

DROP SCHEMA hive.web

Tuning

The following configuration properties may have an impact on connector performance:

hive.assume-canonical-partition-keys

  • Type: Boolean
  • Default value: false
  • Description: Enable optimized metastore partition fetching for non-string partition keys. Setting this property allows to filter non-string partition keys while reading them from hive, based on the assumption that they are stored in canonical (java) format. This is disabled by default as hive allows to use non-canonical format as well (eg. boolean value false may be represented as 0, false, False and more). Used correctly this property may drastically improve read time by reducing number of partition loaded from hive. Setting this property for non-canonical data format may cause erratic behavior.

hive.domain-compaction-threshold

  • Type: Integer (at least 1)
  • Default value: 100
  • Description: Maximum number of ranges/values allowed while reading hive data without compacting it. A higher value will cause more data fragmentation but allow the use of the row skipping feature when reading ORC data. Increasing this value may have a large impact on IN and OR clause performance in scenarios making use of row skipping.

hive.force-local-scheduling

  • Type: Boolean
  • Default value: false
  • Description: Force splits to be scheduled on the same node (ignoring normal node selection procedures) as the Hadoop DataNode process serving the split data. This is useful for installations where Presto is collocated with every DataNode and may increase queries time significantly. The drawback may be that if some data are accessed more often, the utilization of some nodes may be low even if the whole system is heavy loaded. See also node-scheduler.network-topology if less strict constrain is preferred - especially if some nodes are overloaded and other are not fully utilized.

hive.max-initial-split-size

  • Type: String (data size)
  • Default value: hive.max-split-size / 2 (32 MB)
  • Description: This property describes the maximum size of the first hive.max-initial-splits splits created for a query. the logic behind initial splits is described in hive.max-initial-splits. Lower values will increase concurrency for small queries. This property represents the maximum size, as the real size may be lower when the amount of data to read is less than hive.max-initial-split-size (e.g. at the end of a block on a DataNode).

hive.max-initial-splits

  • Type: Integer
  • Default value: 200
  • Description: This property describes how many splits may be initially created for a single query using hive.max-initial-split-size instead of hive.max-split-size. A higher value will force more splits to have a smaller size (hive.max-initial-splits is expected to be smaller than hive.max-split-size), effectively increasing the definition of what is considered a “small query”. The purpose of the smaller split size for the initial splits is to increase concurrency for smaller queries.

hive.max-outstanding-splits

  • Type: Integer (at least 1)
  • Default value: 1000
  • Description: Limit on the nubmer of splits waiting to be served by a split source. After reaching this limit, writers will stop writing new splits until some of hteme are used by workers. Higher values will increase memory usage, but allow IO to be concentrated at one time, which may be faster and increase resource utilization.

hive.max-partitions-per-writers

  • Type: Integer (at least 1)
  • Default value: 100
  • Description: Maximum number of partitions per writer. A query will fail if it requires more partitions per writer than allowed by this property. It can be helpful to have queries beyond the expected maximum partitions to fail to help with error detection. Also it may allow to preactivly avoid out of memory problem.

hive.max-split-iterator-threads

  • Type: Integer (at least 1)
  • Default value: 1000
  • Description: This property describes how many threads may be used to iterate through splits when loading them to the worker nodes. A higher value may increase parallelism, but increased concurrency may cause too much time to be spent on context switching.

hive.max-split-size

  • Type: String (data size)
  • Default value: 64 MB
  • Description: The maximum size of splits created after the initial splits. The logic for initial splits is described in hive.max-initial-splits. A higher value will reduce parallelism. This may be desirable for very large queries and a stable cluster because it allows for more efficient processing of local data without the context switching, synchronization and data collection that result from parallelization. The optimal value should be aligned with the average query size in the system.

hive.metastore.partition-batch-size.max

  • Type: Integer (at least 1)
  • Default value: 100
  • Description: This together with hive.metastore.partition-batch-size.min defines the range of partition sizes read from Hive. The first partition is always of size hive.metastore.partition-batch-size.min and each following partition is two times bigger than previous up to hive.mestastore.partition-batch-size.max (the formula for partition size n is min(hive.metastore.partition-batch-size.max, (2``^``n) * hive.metastore.partition-batch-size.min)). This algorithm allows for live adjustment of partition size according to the processing requirements. If the queries in the system will differ significantly from each other in size, then this range should be extended to better adjust to processing requirements. If the queries in the system will mostly be of the same size, then setting both values to the same maximally tuned value may give a slight edge in processing time.

hive.metastore.partition-batch-size.min

  • Type: Integer (at least 1)
  • Default value: 10
  • Description: See hive.metastore.partition-batch-size.max.

hive.orc.max-buffer-size

  • Type: String (data size)
  • Default value: 8 MB
  • Description: Serves as default value for orc_max_buffer_size session properties defining max size of ORC read operators. Higher value will allow bigger chunks to be processed but will decrease concurrency level.

hive.orc.max-merge-distance

  • Type: String (data size)
  • Default value: 1 MB
  • Description: Serves as the default value for the orc_max_merge_distance session property. Two reads from an ORC file may be merged into a single read if the distance between the requested data ranges in the data source is less than or equal to this value.

hive.orc.stream-buffer-size

  • Type: String (data size)
  • Default value: 8 MB
  • Description: Serves as the default value for the orc_max_buffer_size session property. It defines the maximum size of ORC read operators. A higher value will allow bigger chunks to be processed, but will decrease concurrency.

hive.orc.use-column-names

  • Type: Boolean
  • Default value: false
  • Description: Access ORC columns using names from the file. By default, Hive access columns in ORC files using the order recoded in the Hive metastore. Setting this property allows to use columns names recorded in the ORC file instead.

hive.parquet-optimized-reader.enabled

  • Type: Boolean
  • Default value: false
  • Description: Deprecated Serves as default value for parquet_optimized_reader_enabled session property. Enables number of reader improvements introduced by alternative parquet implementation. The new reader supports vectorized reads, lazy loading, and predicate push down, all of which make the reader more efficient and typically reduces wall clock time for a query. However as the code has changed significantly it may or may not introduce some minor issues, so it can be disabled if some problems with environment are noticed. This property enables/disables all optimizations except predicate push down as it is managed by hive.parquet-predicate-pushdown.enabled property.

hive.parquet-predicate-pushdown.enabled

  • Type: Boolean
  • Default value: false
  • Description: Deprecated Serves as default value for parquet_predicate_pushdown_enabled sesssion property. See hive.parquet-optimized-reader.enabled.

hive.parquet.use-column-names

  • Type: Boolean
  • Default value: false
  • Description: Access Parquet columns using names from the file. By default, columns in Parquet files are accessed by their ordinal position in the Hive metastore. Setting this property allows access by column name recorded in the Parquet file instead.

hive.s3.max-connections

  • Type: Integer (at least 1)
  • Default value: 500
  • Description: The maximum number of connections to S3 that may be open at a time by the S3 driver. A higher value may increase network utilization when a cluster is used on a high speed network. However, a higher values relies more on S3 servers being well configured for high parallelism.

hive.s3.multipart.min-file-size

  • Type: String (data size, at least 16 MB)
  • Default value: 16 MB
  • Description: This property describes how big a file must be to be uploaded to an S3 cluster using the multipart upload feature. Amazon recommends using 100 MB, but a lower value may increase upload parallelism and decrease the data lost/data sent ratio in unstable network conditions.

hive.s3.multipart.min-part-size

  • Type: String (data size, at least 5 MB)
  • Default value: 5 MB
  • Description: Defines the minimum part size for upload parts. Decreasing the minimum part size causes multipart uploads to be split into a larger number of smaller parts. Setting this value too low has a negative effect on transfer speeds, causing extra latency and network communication for each part.

There are also following session properties allowing to control connector behavior on single query basis:

orc_max_buffer_size

orc_max_merge_distance

orc_stream_buffer_size

Hive Connector Limitations

DELETE is only supported if the WHERE clause matches entire partitions.