arrow-bcp 0.9.2

Transfer data between pyarrow and SQL Server using Microsoft's Bulk Copy Program

arrow-bcp

pip install arrow-bcp

When loading data into SQL Server from Python, the common methods are based on the ODBC interface and driver. While this is a tried and tested approach, the performance is not optimal. Reasons for this include:

• minimal logging cannot be used by regular insert statements unless they use table-valued parameters, which are cumbersome to use.
• General overhead caused by the ODBC specification. Most databases other than SQL Server provide their own connector in addition to an ODBC driver for this reason

Microsoft's Bulk Copy Program (bcp) command-line utility seems to use additional techniques beyond the ODBC specification and achieves better performance.

The purpose of this library is to (de-)serialize pyarrow dataframes into SQL Server's native format so that it can be understood and ingested by bcp, or equivalently, SQL Server's BULK INSERT statement.

Some wrappers around the bcp CLI are provided as well to simplify interaction with it.

For example usage including a benchmark for inserts, see the notebooks in the examples directory.

Prerequisites

• An ODBC Driver Manager. On Linux/Mac, you may need to install unixODBC. On Windows, a driver manager is preinstalled.
• The ODBC Driver for SQL Server, which also bundles the bcp executable (Link).
• If the ODBC Driver for SQL Server was not installed using Microsoft's install script (e.g., by copy-pasting the binaries onto a system), you may need to configure odbc.ini (located in /etc/ on Linux) such that "ODBC Driver 17 for SQL Server," or whichever version you installed, points to the ODBC driver executable. If bcp is not on the PATH, you will also need to specify its location using arrow_bcp.set_bcp_executable(path).

Type mapping

Arrow to SQL Server

Arrow	SQL Server
utf8	varchar(max) (utf8)
binary	varbinary(max)
fixedbinary	varbinary(max)
decimal128(p, s)	decimal(p, s)
boolean	bit
int8	smallint
uint8	tinyint
int16	smallint
uint16	int
int32	int
uint32	bigint
int64	bigint
uint64	decimal(20, 0)
float16	real
float32	real
float64	float
date32	date
date64	datetime2
timestamp(s/ms/us/ns)	datetime2
time32(s/ms)	time
time64(us/ns)	time
timestamp(s/ms/us/ns+tz)	datetimeoffset
null	varbinary(max)

SQL Server to Arrow

SQL Server	Arrow
(n)(var)char	utf8
(n)text	utf8
variant	utf8
(var)binary	binary
user defined type	binary
uniqueidentifier	fixedbinary(16)
decimal(p, s)	decimal128(p, s) or null
numeric(p, s)	decimal128(p, s) or null
(small)money	decimal128
bit	boolean
tinyint	uint8
smallint	int16
int	int32
bigint	int64
real	float32
float	float64
date	date32
(small)datetime	timestamp(us)
datetime2	timestamp(us)
time	time64(ns)
datetimeoffset	timestamp(us+tz) or null

Limitations

Reading data from SQL Server

Decimals / timestamps with timezones

In both Arrow and SQL Server's format, decimal size/precision and timezone offset are part of the column's datatype. However, bcp does not provide this information as metadata and instead specifies it separately for each cell (even though all cells must match). As a consequence, the datatype of an Arrow decimal/timezone column is only known once the first non-null cell is read. If all cells are null, we don't get this information, so the entire datatype is set to null as a workaround. This may lead to issues when another datatype is expected.

Datetime2 / datetimeoffset

Both of these datatypes store time with an accuracy of 100ns. However Arrow only offers datatype with accuracy 1ns or 1000ns (1us) and both have tradeoffs. The 1ns datatype cannot represent larger values like '9999-12-31' and the 1000ns datatype truncates the last digit, i.e. '9999-12-31 11:11:11.1234567' turns to '9999-12-31 11:11:11.1234560'. This library chooses to truncate.

Acknowledgements

Thanks to Adam Serafini for figuring out the build process for Zig extensions and providing an example module. He explains that it was quite a journey in his talk, which I appreciate, as even just figuring out that targeting x86_64-windows instead of x86_64-windows-msvc works better for Windows compilation took me several hours.