/elephant-bird

Twitter's collection of LZO and Protocol Buffer-related Hadoop, Pig, and HBase code.

Primary LanguageJavaApache License 2.0Apache-2.0

Elephant Bird

Twitter's library of LZO and/or Protocol Buffer-related Hadoop InputFormats, OutputFormats, Writables, Pig LoadFuncs, HBase miscellanea, etc. The majority of these are in production at Twitter running over data every day.

To Use

  1. git clone
  2. ant
  3. check out javadoc, etc.
  4. build examples: cd examples && ant;

Note: for any of the LZO-based code, make sure that the native LZO libraries are on your java.library.path. Generally this is done by setting JAVA_LIBRARY_PATH in pig-env.sh or hadoop-env.sh. You can also add lines like 

PIG_OPTS=-Djava.library.path=/path/to/my/libgplcompression/dir
to pig-env.sh. See the instructions for Hadoop-LZO for more details.

There are a few simple examples that use the input formats. Note how the protocol buffer-based formats work, and also note that the examples build file uses the custom codegen stuff. See below for more about that.

Building Without Protocol Buffers

If you don't want to build elephant-bird with protobuf support or you don't have protobuf >= 2.3 available (fedora for instance only provides 2.2), you can have ant exclude all classes that rely on protobuf by using ant noproto target

for the default target this would be: 

ant noproto release-jar

License

Apache licensed.

Contents

LZO-based Hadoop Input Formats
  • JSON data
  • Line-based data (TextInputFormat but for LZO; also available in deprecated 0.18 format)
  • W3C logs
  • Serialized protocol buffers in one of three flavors
    • Block-based (via codegen, see below; also available in deprecated 0.18 format)
    • Block-based, into generic bytes.
    • Line-based, base64 encoded (via codegen, see below)
LZO-based Hadoop Writables
  • Protocol buffer writables (via codegen, see below)
LZO-based Hadoop OutputFormats
  • Serialized protocol buffers in one of two flavors
    • Block-based (via codegen, see below)
    • Line-based, base64 encoded (via codegen, see below)
LZO-based LoadFuncs for Pig
  • JSON data
  • Regex-based loaders
  • Sampling loaders for LZO
  • Text loaders (TextLoader equivalent but for LZO)
  • Tokenized loaders (PigStorage equivalent but for LZO)
  • W3C logs
  • Serialized protocol buffers in one of two flavors
    • Block-based (via codegen, see below)
    • Line-based, base64 encoded (via codegen, see below)
LZO-based StoreFuncs for Pig
  • Tokenized storage (PigStorage equivalent but for LZO)
  • Serialized protocol buffers in one flavor
    • Line-based, base64 encoded (via codegen, see below)
Utilities
  • Counters in Pig
  • Protocol buffer utilities
  • Conversions from protocol buffers to pig scripts
  • Reading and writing block-based protocol buffer format (see ProtobufBlockWriter)

Protobuf Codegen?

Yes. Most of the work with protobufs can be templatized via a <M extends Message> (where Message means com.google.protobuf.Message) but Hadoop works mainly via reflection. Java type erasure prevents templated types from being instantiated properly via reflection, so our solution is to write the templatized class, and then generate code for derived classes that do little more than instantiate the type parameter. This causes a slight proliferation of classes, but it only needs to be done during the build phase (don't check in generated code!) so in practice it isn't an issue. It turns out this can be done for Hadoop, Pig, HBase, etc, and you can easily add your own classes to it. The model we use here is file-based: essentially, you can configure it such that you generate some set of derived protobuf-related classes for each protocol buffer defined in a given file. This distribution ships with classes

In protobuf 2.3, Google introduced the notion of a protocol buffer plugin that lets you hook in to their code generation elegantly, with all the parsed metadata available. We use this in com.twitter.elephantbird.proto.HadoopProtoCodeGenerator to generate code for each protocol buffer. The HadoopProtoCodeGenerator expects as a first argument a yml file consisting of keys and lists of classnames. For each protocol buffer file read in (say from my_file.proto), it looks up the basename (my_file) in the yml file.
If a corresponding list exists, it expects each element is a classname of a class deriving from com.twitter.elephantbird.proto.ProtoCodeGenerator. These classes implement a method to set the filename, and a method to set the generated code contents of the file. You can add your own by creating such a derived class and including it in the list of classnames for the protocol buffer file key. That is, if you want to apply the code generators in com.twitter.elephantbird.proto.codegen.ProtobufWritableGenerator and com.twitter.elephantbird.proto.codegen.LzoProtobufBytesToPigTupleGenerator to every protobuf in the file my_file.proto, then your config file should have a section that looks like 

my_file:


    

  • com.twitter.elephantbird.proto.codegen.ProtobufWritableGenerator
    • 

  • com.twitter.elephantbird.proto.codegen.LzoProtobufBytesToPigTupleGenerator

There are examples in the examples subdirectory showing how to integrate this code generation into a build, both for generating Java files pre-jar and for generating other types of files from protocol buffer definitions post-compile (there are examples that do this to generate Pig loaders for a set of protocol buffers).

Commit Back!

Bug fixes, features, and documentation improvements are welcome! Please fork and send me a pull request on github, and I will do my best to keep up. If you make major changes, add yourself to the contributors list below.

Contributors