This project is an enhancement of the Python database module for ODBC that implements the Python DB API 2.0 specification. You can see the original project here:


The enhancements are documented in this file. For general info about the pyodbc package, please refer to the original project documentation.

This module enhancement requires:

  • Python 2.4 or greater
  • ODBC 3.0 or greater
  • NumPy 1.5 or greater (1.7 is required for datetime64 support)

The enhancements in this module consist mainly in the addition of some new methods for fetching the data after a query and put it in a variety of NumPy containers.

Using NumPy as data containers instead of the classical list of tuples has a couple of advantages:

1) The NumPy container is much more compact, and hence, it requires much less memory, than the original approach.

2) As a NumPy container can hold arbitrarily large arrays, it requires much less object creation than the original approach (one Python object per datum retrieved).

This means that this enhancements will allow to fetch data out of relational databases in a much faster way, while consuming significantly less resources.

API additions


  • pyodbc.npversion The version for the NumPy additions


Cursor.fetchdictarray (size=cursor.arraysize)

This is similar to the original Cursor.fetchmany(size), but the data is returned in a dictionary where the keys are the names of the columns and the values are NumPy containers.

For example, it a SELECT is returning 3 columns with names ‘a’, ‘b’ and ‘c’ and types varchar(10), integer and timestamp, the returned object will be something similar to:

{'a': array([...], dtype='S11'),
 'b': array([...], dtype=int32),
 'c': array([...], dtype=datetime64[us])}

Note that the varchar(10) type is translated automatically to a string type of 11 elements (‘S11’). This is because the ODBC driver needs one additional space to put the trailing ‘0’ in strings, and NumPy needs to provide the room for this.

Also, it is important to stress that all the timestamp types are translated into a NumPy datetime64 type with a resolution of microseconds by default.

Cursor.fetchsarray (size=cursor.arraysize)

This is similar to the original Cursor.fetchmany(size), but the data is returned in a NumPy structured array, where the name and type of the fields matches to those resulting from the SELECT.

Here it is an example of the output for the SELECT above:

      dtype=[('a', '|S11'), ('b', '<i4'), ('c', ('<M8[us]', {}))])

Note that, due to efficiency considerations, this method is calling the fetchdictarray() behind the scenes, and then doing a conversion to get an structured array. So, in general, this is a bit slower than its fetchdictarray() counterpart.

Data types supported

The new methods listed above have support for a subset of the standard ODBC. In particular:

  • String support (SQL_VARCHAR) is supported.
  • Numerical types, be them integers or floats (single and double precision) are fully supported. Here it is the complete list: SQL_INTEGER, SQL_TINYINT, SQL_SMALLINT, SQL_FLOAT and SQL_DOUBLE.
  • Dates, times, and timestamps are mapped to the datetime64 and timedelta NumPy types. The list of supported data types are: SQL_DATE, SQL_TIME and SQL_TIMESTAMP,
  • Binary data is not supported yet.
  • Unicode strings are not supported yet.

NULL values

As there is not (yet) a definitive support for missing values (NA) in NumPy, this module represents NA data as particular values depending on the data type. Here it is the current table of the particular values:

int8: -128 (-2**7)
uint8: 255 (2**8-1)
int16: -32768 (-2**15)
uint16: 65535 (2**16-1)
int32: -2147483648 (-2**31)
uint32: 4294967295 (2**32-1)
int64: -9223372036854775808 (-2**63)
uint64: 18446744073709551615 (2**64-1)
float32: NaN
float64: NaN
datetime64: NaT
timedelta64: NaT (or -2**63)
string: 'NA'

Improvements for 1.1 release

  • The rowcount is not trusted anymore for the fetchdict() and fetchsarray() methods. Now the NumPy containers are built incrementally, using realloc for a better use of resources.
  • The Python interpreter does not exit anymore when fetching an exotic datatype not supported by NumPy.
  • The docsctrings for fetchdict() and fetchsarray() have been improved.