Table of Contents

Gobblin has the ability to register the ingested/compacted data in Hive. This allows registering data in Hive immediately after data is published at the destination, offering much lower latency compared to doing data ingestion and Hive registration separately.

How Hive Registration Works in Gobblin

Hive registration is done in HiveRegister. After the data is published, the publisher or compaction runner will create an instance of HiveRegister, and for each path that should be registered in Hive, the publisher or compaction runner will use a specific HiveRegistrationPolicy to create a list of HiveSpecs for the path. It creates a list of HiveSpecs rather than a single HiveSpec for each path, so that the same path can be registered in multiple tables or databases.

HiveSpec

A HiveSpec specifies how a path should be registered in Hive, i.e., which database, which table, which partition should the path be registered. An example is SimpleHiveSpec.

HiveRegistrationPolicy

HiveRegistrationPolicy is responsible for generating HiveSpecs given a path. For example, if you want paths ending with a date (e.g., /(something)/2016/05/22) to be registered in the corresponding daily partition (e.g., daily-2016-05-22), you can create an implementation of HiveRegistrationPolicy that contains the logic of converting such a path into a Hive partition.

An example is HiveRegistrationPolicyBase, which provides base implementation for getting database names and table names for a path:

  • A database/table name can be specified explicitly in hive.database.name or hive.table.name.
  • Alternatively, a database/table regex can be provided in hive.database.regex or hive.table.regex. The regex will be matched against the path to be registered, and if they match, the first group is considered the database/table name.
  • It is possible to register a path to multiple databases or tables by specifying additional.hive.database.names and additional.hive.table.names. If multiple databases and tables are specified, the path will be registered to the cross product.
  • If the provided/derived Hive database/table names are invalid, they are sanitized into a valid name. A database/table name is valid if it starts with an alphanumeric character, contains only alphanumeric characters and _, and is not composed of numbers only.

One should in general extend HiveRegistrationPolicyBase when implementing a new HiveRegistrationPolicy.

HiveSerDeManager

If the data to be registered is in a format other than plain text (CSV, TSV, etc.), you often need to use a SerDe and specify some SerDe properties including the type of SerDe, input format, output format, schema, etc. This is done in HiveSerDeManager, which is part of a HiveRegistrationUnit (i.e., HiveTable or HivePartition). An example is HiveAvroSerDeManager.

Predicate and Activity

One or more Predicates can be attached to a HiveSpec. If a HiveSpec contains Predicates, unless Predicates return true, the HiveSpec will not be registered. This is useful in cases where, for example, one only wants to register a path in Hive if a particular Hive table or partition doesn't already exist. An example is TableNotExistPredicate.

One or more Activitys can be attached to a HiveSpec. There are two types of activities: pre-activities and post-activities, which will be executed before and after a HiveSpec is registered, respectively. This is useful, for example, when you need to drop/alter a table/partition before or after a path is registered. An example is DropTableActivity.

How to Use Hive Registration in Your Gobblin Job

First, is to implement a HiveRegistrationPolicy (or reuse an existing one), then specify its class name in config property hive.registration.policy.

Then, specify the appropriate table/partition properties in hive.table.partition.props, storage descriptor properties in hive.storage.props, and SerDe properties in hive.serde.props. Some SerDe properties are usually dynamic (e.g., schema), which are added in the HiveSerDeManager.

Example table/partition properties are "owner" and "retention", example storage descriptor properties are "location", "compressed", "numBuckets", example SerDe properties are "serializationLib", "avro.schema.url".

If you are running a Gobblin ingestion job:

  • If data is published in the job (which is the default case), use a job-level data publisher that can perform Hive registration, such as BaseDataPublisherWithHiveRegistration. If you need to do Hive registration with a different publisher than BaseDataPublisher, you will need to extend that publisher to do Hive registration, which will be similar as how BaseDataPublisher is extended into BaseDataPublisherWithHiveRegistration.
  • If data is published in the tasks, use HiveRegistrationPublisher as the job-level data publisher. This publisher does not publish any data; it only does Hive registration.

If you are running a Gobblin compaction job: add HiveRegistrationCompactorListener to the list of compaction listeners by adding the class name to property compaction.listeners.

Hive Registration Config Properties

Property Name Semantics
hive.registration.policy Class name which implements HiveRegistrationPolicy
hive.row.format Either AVRO, or the class name which implements HiveSerDeManager
hive.database.name Hive database name
hive.database.regex Hive database regex
hive.database.name.prefix Hive database name prefix
hive.database.name.suffix Hive database name suffix
additional.hive.database.names Additional Hive database names
hive.table.name Hive table name
hive.table.regex Hive table regex
hive.table.name.prefix Hive table name prefix
hive.table.name.suffix Hive table name suffix
additional.hive.table.names Additional Hive table names
hive.register.threads Thread pool size used for Hive registration
hive.db.root.dir The root dir of Hive db
hive.table.partition.props Table/partition properties
hive.storage.props Storage descriptor properties
hive.serde.props SerDe properties
hive.registration.fs.uri File system URI for Hive registration
hive.upstream.data.attr.names Attributes to describe upstream data source as Hive Metadata