Overview - Working with Avro from Hive
The AvroSerde allows users to read or write Avro data as Hive tables. The AvroSerde's bullet points:
- Infers the schema of the Hive table from the Avro schema.
- Reads all Avro files within a table against a specified schema, taking advantage of Avro's backwards compatibility abilities
- Supports arbitrarily nested schemas.
- Translates all Avro data types into equivalent Hive types. Most types map exactly, but some Avro types don't exist in Hive and are automatically converted by the AvroSerde.
- Understands compressed Avro files.
- Transparently converts the Avro idiom of handling nullable types as Union[T, null] into just T and returns null when appropriate.
- Writes any Hive table to Avro files.
- Has worked reliably against our most convoluted Avro schemas in our ETL process.
The AvroSerde has been built and tested against Hive 0.9.1 and Avro 1.5.
Avro to Hive type conversion
While most Avro types convert directly to equivalent Hive types, there are some which do not exist in Hive and are converted to reasonable equivalents. Also, the AvroSerde special cases unions of null and another type, as described below:
Becomes Hive type
Bytes are converted to Array[smallint] prior to Hive 0.12.0
Unions of [T, null] transparently convert to nullable T, other types translate directly to Hive's unions of those types. However, unions were introduced in Hive 7 and are not currently able to be used in where/group-by statements. They are essentially look-at-only. Because the AvroSerde transparently converts [T,null], to nullable T, this limitation only applies to unions of multiple types or unions not of a single type and null.
Hive has no concept of enums
Fixeds are converted to Array[smallint] prior to Hive 0.12.0
Creating Avro-backed Hive tables
To create a the Avro-backed table, specify the serde as org.apache.hadoop.hive.serde2.avro.AvroSerDe, specify the inputformat as org.apache.hadoop.hive.ql.io.avro.AvroContainerInputFormat, and the outputformat as org.apache.hadoop.hive.ql.io.avro.AvroContainerOutputFormat. Also provide a location from which the AvroSerde will pull the most current schema for the table. For example:
In this example we're pulling the source-of-truth reader schema from a webserver. Other options for providing the schema are described below.
Add the Avro files to the database (or create an external table) using standard Hive operations(http://wiki.apache.org/hadoop/Hive/LanguageManual/DML).
This table might result in a description as below:
At this point, the Avro-backed table can be worked with in Hive like any other table.
Writing tables to Avro files
The AvroSerde can serialize any Hive table to Avro files. This makes it effectively an any-Hive-type to Avro converter. In order to write a table to an Avro file, you must first create an appropriate Avro schema. Create as select type statements are not currently supported. Types translate as detailed in the table above. For types that do not translate directly, there are a few items to keep in mind:
- Types that may be null must be defined as a union of that type and Null within Avro. A null in a field that is not so defined with result in an exception during the save. No changes need be made to the Hive schema to support this, as all fields in Hive can be null.
- Avro Bytes type should be defined in Hive as lists of tiny ints. the AvroSerde will convert these to Bytes during the saving process.
- Avro Fixed type should be defined in Hive as lists of tiny ints. the AvroSerde will convert these to Fixed during the saving process.
- Avro Enum type should be defined in Hive as strings, since Hive doesn't have a concept of enums. Ensure that only valid enum values are present in the table - trying to save a non-defined enum will result in an exception.
Consider the following Hive table, which coincidentally covers all types of Hive data types, making it a good example:
To save this table as an Avro file, create an equivalent Avro schema (the namespace and actual name of the record are not important):
If the table were backed by a csv such as:
why hello there
2:Time to go home
one can write it out to Avro with:
The files that are written by the Hive job are valid Avro files, however, MapReduce doesn't add the standard .avro extension. If you copy these files out, you'll likely want to rename them with .avro.
Hive is very forgiving about types: it will attempt to store whatever value matches the provided column in the equivalent column position in the new table. No matching is done on column names, for instance. Therefore, it is incumbent on the query writer to make sure the the target column types are correct. If they are not, Avro may accept the type or it may throw an exception, this is dependent on the particular combination of types.
Specifying the Avro schema for a table
There are three ways to provide the reader schema for an Avro table, all of which involve parameters to the serde. As the schema involves, one can update these values by updating the parameters in the table.
Specifies a url to access the schema from. For http schemas, this works for testing and small-scale clusters, but as the schema will be accessed at least once from each task in the job, this can quickly turn the job into a DDOS attack against the URL provider (a web server, for instance). Use caution when using this parameter for anything other than testing.
The schema can also point to a location on HDFS, for instance: hdfs://your-nn:9000/path/to/avsc/file. the AvroSerde will then read the file from HDFS, which should provide resiliency against many reads at once. Note that the serde will read this file from every mapper, so it's a good idea to turn the replication of the schema file to a high value to provide good locality for the readers. The schema file itself should be relatively small, so this does not add a significant amount of overhead to the process.
Use schema.literal and embed the schema in the create statement
One can embed the schema directly into the create statement. This works if the schema doesn't have any single quotes (or they are appropriately escaped), as Hive uses this to define the parameter value. For instance:
Note that the value is enclosed in single quotes and just pasted into the create statement.
Use avro.schema.literal and pass the schema into the script
Hive can do simple variable substitution and one can pass the schema embedded in a variable to the script. Note that to do this, the schema must be completely escaped (carriage returns converted to \n, tabs to \t, quotes escaped, etc). An example:
To execute this script file, assuming $SCHEMA has been defined to be the escaped schema value:
Note that $SCHEMA is interpolated into the quotes to correctly handle spaces within the schema.
Use none to ignore either avro.schema.literal or avro.schema.url
Hive does not provide an easy way to unset or remove a property. If you wish to switch from using url or schema to the other, set the to-be-ignored value to none and the AvroSerde will treat it as if it were not set.
If something goes wrong
Hive tends to swallow exceptions from the AvroSerde that occur before job submission. To force Hive to be more verbose, it can be started with *hive -hiveconf hive.root.logger=INFO,console*, which will spit orders of magnitude more information to the console and will likely include any information the AvroSerde is trying to get you about what went wrong. If the AvroSerde encounters an error during MapReduce, the stack trace will be provided in the failed task log, which can be examined from the JobTracker's web interface. the AvroSerde only emits the AvroSerdeException; look for these. Please include these in any bug reports. The most common is expected to be exceptions while attempting to serializing an incompatible type from what Avro is expecting.
- Why do I get error-error-error-error-error-error-error and a message to check avro.schema.literal and avro.schema.url when describing a table or running a query against a table?
The AvroSerde returns this message when it has trouble finding or parsing the schema provided by either the avro.schema.literal or avro.avro.schema.url value. It is unable to be more specific because Hive expects all calls to the serde config methods to be successful, meaning we are unable to return an actual exception. By signaling an error via this message, the table is left in a good state and the incorrect value can be corrected with a call to alter table T set TBLPROPERTIES.