Skip to content

Latest commit

 

History

History
139 lines (89 loc) · 7.23 KB

README.rst

File metadata and controls

139 lines (89 loc) · 7.23 KB
Raw

ibmdbpy

Accelerating Python Analytics by In-Database Processing

The ibmdbpy project provides a Python interface for data manipulation and access to in-database algorithms in IBM dashDB and IBM DB2. It accelerates Python analytics by seamlessly pushing operations written in Python into the underlying database for execution, thereby benefitting from in-database performance-enhancing features, such as columnar storage and parallel processing.

IBM dashDB is a database management system available on IBM Bluemix, the cloud application development and analytics platform powered by IBM. The ibmdbpy project can be used by Python developers with very little additional knowledge, because it copies the well-known interface of the Pandas library for data manipulation and the Scikit-learn library for the use of machine learning algorithms.

The ibmdbpy project is compatible with Python releases 2.7 up to 3.4 and can be connected to dashDB or DB2 instances via ODBC or JDBC.

The project is still at an early stage and many of its features are still in development. However, several experiments have already demonstrated that it provides significant performance advantages when operating on medium or large amounts of data, that is, on tables of 1 million rows or more.

The latest version of ibmdbpy is available on the Python Package Index and Github.

How ibmdbpy works

The ibmdbpy project translates Pandas-like syntax into SQL and uses a middleware API (pypyodbc/JayDeBeApi) to send it to an ODBC or JDBC-connected database for execution. The results are fetched and formatted into the corresponding data structure, for example, a Pandas.Dataframe or a Pandas.Series.

The following scenario illustrates how ibmdbpy works.

Assuming that all ODBC connection parameters are correctly set, issue the following statements to connect to a database (in this case, a dashDB instance named DASHDB) via ODBC:

>>> from ibmdbpy import IdaDataBase, IdaDataFrame
>>> idadb = IdaDataBase('DASHDB')

A few sample data sets are included in ibmdbpy for you to experiment. We can firstly load the well-known IRIS table into this dashDB instance.

>>> from ibmdbpy.sampledata import iris
>>> idadb.as_idadataframe(iris, "IRIS")
<ibmdbpy.frame.IdaDataFrame at 0x7ad77f0>

Next, we can create an IDA data frame that points to the table we just uploaded. Let’s use that one:

>>> idadf = IdaDataFrame(idadb, 'IRIS')

Note that to create an IDA data frame using the IdaDataFrame object, we need to specify our previously opened IdaDataBase object, because it holds the connection.

Now let us compute the correlation matrix:

>>> idadf.corr()

In the background, ibmdbpy looks for numerical columns in the table and builds an SQL request that returns the correlation between each pair of columns. Here is the SQL request that was executed for this example:

SELECT CORRELATION("sepal_length","sepal_width"),
CORRELATION("sepal_length","petal_length"),
CORRELATION("sepal_length","petal_width"),
CORRELATION("sepal_width","petal_length"),
CORRELATION("sepal_width","petal_width"),
CORRELATION("petal_length","petal_width")
FROM IRIS

The result fetched by ibmdbpy is a tuple containing all values of the matrix. This tuple is formatted back into a Pandas.DataFrame and then returned:

              sepal_length  sepal_width  petal_length  petal_width
sepal_length      1.000000    -0.117570      0.871754     0.817941
sepal_width      -0.117570     1.000000     -0.428440    -0.366126
petal_length      0.871754    -0.428440      1.000000     0.962865
petal_width       0.817941    -0.366126      0.962865     1.000000

Et voilà !

How the geospatial functions work

The ibmdbpy package now supports geospatial functions! It provides a Python interface for data manipulation and access to in-database algorithms in IBM dashDB and IBM DB2 Spatial Extender. It identifies the geometry column for spatial tables and enables the user to perform spatial queries based upon this column. The results are fetched and formatted into the corresponding data structure, for example, an IdaGeoDataframe.

The following scenario illustrates how spatial functions work.

We can create an IDA geo data frame that points to a sample table in dashDB:

>>> from ibmdbpy import IdaDataBase, IdaGeoDataFrame
>>> idadb = IdaDataBase('DASHDB')
>>> idadf = IdaGeoDataFrame(idadb, 'SAMPLES.GEO_COUNTY')

Note that to create an IdaGeoDataframe using the IdaDataFrame object, we need to specify our previously opened IdaDataBase object, because it holds the connection.

Now let us compute the area of the counties in the GEO_COUNTY table. The result of the area will be stored as a new column 'area' in the IdaGeoDataFrame:

>>> idadf['area'] = idadf.area(colx = 'SHAPE')
    OBJECTID    NAME         SHAPE                                              area
    1           Wilbarger    MULTIPOLYGON (((-99.4756582604 33.8340108094, ...  0.247254
    2           Austin       MULTIPOLYGON (((-96.6219873342 30.0442882117, ...  0.162639
    3           Logan        MULTIPOLYGON (((-99.4497297204 46.6316377481, ...  0.306589
    4           La Plata     MULTIPOLYGON (((-107.4817473750 37.0000108736,...  0.447591
    5           Randolph     MULTIPOLYGON (((-91.2589262966 36.2578866492, ...  0.170844

In the background, ibmdbpy looks for geometry columns in the table and builds an SQL request that returns the area of each geometry. Here is the SQL request that was executed for this example:

SELECT t.*,db2gse.ST_Area(t.SHAPE) as area
FROM SAMPLES.GEO_COUNTY t;

That's as simple as that!

Feature Selection

Ibmdbpy provides a range of functions to support efficient in-database feature selection, e.g. to estimate the relevance of attributes with respect to a particular target. Functions and documentation can be found in the submodule ibmdbpy.feature_selection.

Project Roadmap

  • Full test coverage (a basic coverage is already provided)
  • Add more functions and improve what already exists
  • Add wrappers for several ML-Algorithms (Linear regression, Sequential patterns...)

A more detailed roadmap is available on Github, in the ROADMAP.txt file

Contributors

The ibmdbpy project was initiated in April 2015 at IBM Deutschland Reasearch & Development, Böblingen. Here is the list of the persons who contributed to the project, in the chronological order of their contribution:

  • Edouard Fouché (core)
  • Michael Wurst (core)
  • William Moore (documentation)
  • Craig Blaha (documentation)
  • Rafael Rodriguez Morales (geospatial extension, core)
  • Avipsa Roy (geospatial extension)
  • Nicole Schoen (core)

How to contribute

You want to contribute? That's great! There are many things you can do.

If you are a member of the ibmdbanalytics group, you can create branchs and merge them to master. Otherwise, you can fork the project and do a pull request. You are very welcome to contribute to the code and to the documentation.

There are many ways to contribute. If you found bugs and have improvement ideas or need some new specific features, please open a ticket! We do care about it.