This project brings in-DBMS data analytics to Impala. This leverages previous work done by two projects:

• MADlib (http://madlib.net/)

• Bismarck (http://hazy.cs.wisc.edu/hazy/victor/bismarck/)

Each of these projects use User Defined Aggregates (UDAs) to train analytic models using an existing DBMS's data management and processing ability.

• eigen3 (yum install eigen3-devel)

• boost>=1.54.0 (manual install on CentOS 6)

• impala-udf-devel (yum-installable with Cloudera repo)

make
python python/deploy.py <relevant options>

This is a fork of MADlib 1.0 which has been modified for use with Impala. The specific changes were:

• madlib/test with tests for the new code

• madlib/Makefile to make the tests

• madlib/src/ports/metaport which is a modified MADlib backend for main memory

To run the example SVM,:

1. Create database toysvm

2. Run make. If it fails with an error like lib/libsvm.so: undefined reference to 'impala_udf::FunctionContext::Allocate(int)', it's ok.

3. Register the UDFs with a database (without re-making the binaries), execute: python python/deploy.py -p -o /path/for/libs toysvm

4. Create a synthetic table of examples in the database toysvm with the table toy: python python/gen_classify_data.py toysvm toy

5. python python/impala_svm.py lbl e0 e1 e2 --db toysvm --table toy -e 1

6. impala-shell -q 'use toysvm; select iter, printarray(decodearray(model)) from history;'

Also see example usage in the impyla repo.