This guide explains all of the elements needed to successfully develop and plug in a new MADlib® module.
...
Install MADlib by following the steps in the Installation Guide for MADlib or use the Docker image instructions below.
MADlib source code is organized such that the core logic of a machine learning or statistical module is located in a common location, and the database-port specific code is located in a ports
folder. Since all currently supported databases are based on Postgres, the postgres
port contains all the port-specific files, with greenplum
and hawq
inheriting from it. Before proceeding with this guide, it is recommended that you familiarize yourself with the MADlib module anatomy.Implementation details for existing MADlib algorithms are described in the algorithm technical design document which is a handy reference.is recommended that you familiarize yourself with the MADlib module anatomy.
Anchor | ||||
---|---|---|---|---|
|
We provide a Docker image with necessary dependencies required to compile and test MADlib on PostgreSQL 9.6. You can view the dependency docker file at ./tool/docker/base/Dockerfile_postgres_9_6
. The image is hosted on docker hub at madlib/postgres_9.6:latest
. Later we will provide a similar Docker image for Greenplum Database.
Some useful commands to use the Docker file:
Code Block | ||
---|---|---|
| ||
## 1) Pull down the `madlib/postgres_9.6:latest` image from docker hub:
docker pull madlib/postgres_9.6:latest
## 2) Launch a container corresponding to the MADlib image, mounting the source code folder to the container:
docker run -d -it --name madlib -v (path to incubator-madlib directory):/incubator-madlib/ madlib/postgres_9.6
where incubator-madlib is the directory where the MADlib source code resides.
############################################## * WARNING * ##################################################
# Please be aware that when mounting a volume as shown above, any changes you make in the "incubator-madlib"
# folder inside the Docker container will be reflected on your local disk (and vice versa). This means that
# deleting data in the mounted volume from a Docker container will delete the data from your local disk also.
#############################################################################################################
## 3) When the container is up, connect to it and build MADlib:
docker exec -it madlib bash
mkdir /incubator-madlib/build-docker
cd /incubator-madlib/build-docker
cmake ..
make
make doc
make install
## 4) Install MADlib:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres install
## 5) Several other madpack commands can now be run:
# Run install check, on all modules:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres install-check
# Run install check, on a specific module, say svm:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres install-check -t svm
# Run dev check, on all modules (more comprehensive than install check):
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres dev-check
# Run dev check, on a specific module, say svm:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres dev-check -t svm
# Reinstall MADlib:
src/bin/madpack -p postgres -c postgres/postgres@localhost:5432/postgres reinstall
## 6) Kill and remove containers (after exiting the container):
docker kill madlib
docker rm madlib |
Anchor | ||||
---|---|---|---|---|
|
...
- Register the module.
- Define the SQL functions.
- Implement the functions in C++.Register the C++ header files
- Register the C++ header files.
The files for this exercise can be found in the hello world folder of the source code repository.
1. Register the module
Add the following line to the file called Modules.yml
under ./src/config/
...
At the beginning of file avg_var.sql_in
the command m4_include
('SQLCommon.m4') ` is necessary to run the m4 macro processor. M4 is used to add platform-specific commands in the SQL definitions and is run while deploying MADlib to the database.
...
Code Block | ||
---|---|---|
| ||
/** * @brief Update state with a new data point */ template <class OtherHandle> AvgVarTransitionState &operator+=(const double x){ double diff = (x - avg); double normalizer = static_cast<double>(numRows + 1); // online update mean this.avg->avg += diff / normalizer; // online update variance double new_diff = (x - avg); double a = static_cast<double>(state.numRows) / normalizer; this.var->var = (var * a) + (diff * new_diff) / normalizer; } /** * @brief Merge with another State object * * We update mean and variance in a online fashion * to avoid intermediate large sum. */ template <class OtherHandle> AvgVarTransitionState &operator+=( const AvgVarTransitionState<OtherHandle> &inOtherState) { if (mStorage.size() != inOtherState.mStorage.size()) throw std::logic_error("Internal error: Incompatible transition " "states"); double avg_ = inOtherState.avg; double var_ = inOtherState.var; uint16uint64_t numRows_ = static_cast<uint16cast<uint64_t>(inOtherState.numRows); double totalNumRows = static_cast<double>(numRows + numRows_); double p = static_cast<double>(numRows) / totalNumRows; double p_ = static_cast<double>(numRows_) / totalNumRows; double totalAvg = avg * p + avg_ * p_; double a = avg - totalAvg; double a_ = avg_ - totalAvg; numRows += numRows_; var = p * var + p_ * var_ + p * a * a + p_ * a_ * a_; avg = totalAvg; return *this; } |
...
Now let's run an example using the new module. First, rebuild and reinstall MADLib according to the instructions from Installation Guide. We use the patients
dataset from the MADlib Quick Start Guide for Users for testing purposes. From the psql
terminal, the result below shows that half of the 20 patients have had second heart attacks within 1 year (yes = 1):
Code Block | ||
---|---|---|
| ||
SELECT madlib.avg_var(second_attack) FROM patients; -- ************ -- -- Result -- -- ************ -- +-------------------+ | avg_var | |-------------------| | [0.5, 0.25, 20.0] | +-------------------+ -- (average, variance, count) -- |
...
Anchor | ||||
---|---|---|---|---|
|
...
Compared to the steps presented in the last session, here we do not need to modify the Modules.yml
file because we are not creating new module. Another difference is that we create an additional .py_in
python file along with the .sql_in
file. That is where most of the iterative logic will be implemented.
The files for this exercise can be found in the hello world folder of the source code repository. Please note that __init__.py_in
is not included in this folder as an empty file will be sufficient for the purposes of this exercise.
1. Overview
The overall logic is split into three parts. All the UDF and UDA are defined in simple_logistic.sql_in
. The transition
, merge
and final
functions are implemented in C++. Those functions together constitute the UDA called __logregr_simple_step
which takes one step from the current state to decrease the logistic regression objective. And finally in simple_logistic.py_in
the plpy
package is used to implement in python a UDF called logregr_simple_train
which invokes __logregr_simple_step
iteratively until convergence.
...
The example below demonstrates the usage of madlib.logregr_simple_train
on the patients
table we used earlier. The trained classification model is stored in the table called logreg_mdl
and can be viewed using standard SQL query.
Code Block | ||
---|---|---|
| ||
SELECT madlib.logregr_simple_train( 'patients', -- source table 'logreg_mdl', -- output table 'second_attack', -- labels 'ARRAY[1, treatment, trait_anxiety]'); -- features SELECT * FROM logreg_mdl; -- ************ -- -- Result -- -- ************ -- +--------------------------------------------------+------------------+ | coef | log_likelihood | |--------------------------------------------------+------------------| | [-6.27176619714, -0.84168872422, 0.116267554551] | -9.42379 | +--------------------------------------------------+------------------+ |
...