High-level wrapper around BCP for high performance data transfers between pandas and SQL Server. No knowledge of BCP required!! (pronounced BEE-CEE-Pandas)

• Quickstart
• IMPORTANT - Read vs. Write
• Benchmarks
  • to_sql
    • Why not just use the new pandas method='multi'?
  • read_sql
• Requirements
  • Database
  • Python User
• Installation
• Usage
  • Credential/Connection object
  • Recommended Usage
• Known Issues
  • Troubleshooting
    • All quote characters appear in the data
    • All delimiter characters appear in the data
    • Write to database fails due to spaces in columns
• Background
  • Existing Solutions
  • Design and Scope
• Testing
  • Testing Requirements
  • What Is Tested?
  • Testing Implementation
• CI/CD
• Contributing
• GitHub Stars Trend

In [1]: import pandas as pd
   ...: import numpy as np
   ...:
   ...: from bcpandas import SqlCreds, to_sql

In [2]: creds = SqlCreds(
   ...:     'my_server',
   ...:     'my_db',
   ...:     'my_username',
   ...:     'my_password'
   ...: )

In [3]: df = pd.DataFrame(
   ...:         data=np.ndarray(shape=(10, 6), dtype=int),
   ...:         columns=[f"col_{x}" for x in range(6)]
   ...:     )

In [4]: df
Out[4]:
     col_0    col_1    col_2    col_3    col_4    col_5
0  4128860  6029375  3801155  5570652  6619251  7536754
1  4849756  7536751  4456552  7143529  7471201  7012467
2  6029433  6881357  6881390  7274595  6553710  3342433
3  6619228  7733358  6029427  6488162  6357104  6553710
4  7536737  7077980  6422633  7536732  7602281  2949221
5  6357104  7012451  6750305  7536741  7340124  7274610
6  7340141  6226036  7274612  7077999  6881387  6029428
7  6619243  6226041  6881378  6553710  7209065  6029415
8  6881378  6553710  7209065  7536743  7274588  6619248
9  6226030  7209065  6619231  6881380  7274612  3014770

In [5]: to_sql(df, 'my_test_table', creds, index=False, if_exists='replace')

In [6]: df2 = pd.read_sql_table(table_name='my_test_table', con=creds.engine)

In [7]: df2
Out[7]:
     col_0    col_1    col_2    col_3    col_4    col_5
0  4128860  6029375  3801155  5570652  6619251  7536754
1  4849756  7536751  4456552  7143529  7471201  7012467
2  6029433  6881357  6881390  7274595  6553710  3342433
3  6619228  7733358  6029427  6488162  6357104  6553710
4  7536737  7077980  6422633  7536732  7602281  2949221
5  6357104  7012451  6750305  7536741  7340124  7274610
6  7340141  6226036  7274612  7077999  6881387  6029428
7  6619243  6226041  6881378  6553710  7209065  6029415
8  6881378  6553710  7209065  7536743  7274588  6619248
9  6226030  7209065  6619231  6881380  7274612  3014770

The big speedup benefit of bcpandas is in the to_sql function, as the benchmarks below show. However, the bcpandas read_sql function actually performs slower than the pandas equivalent. Therefore, the bcpandas read_sql function was deprecated in v5.0 and has now been removed in v6.0+. To read data from SQL to pandas, use the native pandas method pd.read_sql_table or pd.read_sql_query.

See figures below. All code is in the /benchmarks directory. To run the benchmarks, from the root directory of this repository, run python benchmarks/benchmark.py main --help and fill in the command line options that are presented.

Running this will output

1. PNG image of the graph
2. JSON file of the benchmark data
3. JSON file with the environment details of the machine that was used to generate it

I didn't bother including the pandas non-multiinsert version here because it just takes way too long

1. Because it is still much slower
2. Because you are forced to set the chunksize parameter to a very small number for it to work - generally a bit less then 2100/<number of columns>. This is because SQL Server can only accept up to 2100 parameters in a query. See here and here for more discussion on this, and the recommendation to use a bulk insert tool such as BCP. It seems that SQL Server simply didn't design the regular INSERT statement to support huge amounts of data.

As you can see, pandas native clearly wins here

Any version of Microsoft SQL Server. Can be installed on-prem, in the cloud, on a VM, or one of the Azure versions.

• BCP Utility
• Microsoft ODBC Driver 11, 13, 13.1, or 17 for SQL Server. (Microsoft Docs) See the pyodbc docs for more on different driver versions.
• Python >= 3.8
• pandas >= 1.5
• sqlalchemy >= 1.4
• pyodbc as the supported DBAPI

1. Create creds (see next section)
2. Replace any df.to_sql(...) in your code with bcpandas.to_sql(df, ...)

That's it!

Bcpandas requires a bcpandas.SqlCreds object in order to use it, and also a sqlalchemy.Engine. The user has 2 options when constructing it.

1. Create the bcpandas SqlCreds object with just the minimum attributes needed (server, database, username, password), and bcpandas will create a full Engine object from this. It will use pyodbc, sqlalchemy, and the Microsoft ODBC Driver for SQL Server, and will store it in the .engine attribute.

   In [1]: from bcpandas import SqlCreds
   
   In [2]: creds = SqlCreds('my_server', 'my_db', 'my_username', 'my_password')
   
   In [3]: creds.engine
   Out[3]: Engine(mssql+pyodbc:///?odbc_connect=Driver={ODBC Driver 17 for SQL Server};Server=tcp:my_server,1433;Database=my_db;UID=my_username;PWD=my_password)

2. Pass a full Engine object to the bcpandas SqlCreds object, and bcpandas will attempt to parse out the server, database, username, and password to pass to the command line utilities. If a DSN is used, this will fail.

   (continuing example above)

   In [4]: creds2 = SqlCreds.from_engine(creds.engine)
   
   In [5]: creds2.engine
   Out[5]: Engine(mssql+pyodbc:///?odbc_connect=Driver={ODBC Driver 17 for SQL Server};Server=tcp:my_server,1433;Database=my_db;UID=my_username;PWD=my_password)
   
   In [6]: creds2
   Out[6]: SqlCreds(server='my_server', database='my_db', username='my_username', with_krb_auth=False, engine=Engine(mssql+pyodbc:///?odbc_connect=Driver={ODBC Driver 17 for SQL Server};Server=tcp:my_server,1433;Database=my_db;UID=my_username;PWD=my_password), password=[REDACTED])

built with the help of https://www.tablesgenerator.com/markdown_tables# and https://gist.github.com/rxaviers/7360908

Here are some caveats and limitations of bcpandas.

• Bcpandas has been tested with all ASCII characters 32-127. Unicode characters beyond that range have not been tested.
• An empty string ("") in the dataframe becomes NULL in the SQL database instead of remaining an empty string.
• Because bcpandas first outputs to CSV, it needs to use several specific characters to create the CSV, including a delimiter and a quote character. Bcpandas attempts to use characters that are not present in the dataframe for this, going through the possilbe delimiters and quote characters specified in constants.py. If all possible characters are present in the dataframe and bcpandas cannot find both a delimiter and quote character to use, it will throw an error.
  • The BCP utility does not ignore delimiter characters when surrounded by quotes, unlike CSVs - see here in the Microsoft docs.
• If there is a NaN/Null in the last column of the dataframe it will throw an error. This is due to a BCP issue. See my issue with Microsoft about this here. This doesn't seem to be a problem based on the tests.

If you encounter the error:

bcpandas.constants.BCPandasValueError: Data contains all of the possible quote characters ('"', "'", '`', '~'),
cannot use BCP to import it. Replace one of the possible quote characters in
your data, or use another method besides bcpandas.

And want to still use BCPandas, you will need to pick a quote character and remove all instances of it from the dataframe. Note that you are modifying your data and take care that replacing e.g., the ~ character will not have undesired consequences.

In this case we are looking to remove ~, replacing it with -. Hunt for its presence in a column:

my_df['some_text_column'].str.contains('~').sum()

If that returns a value greater than zero, you can perform replacement in that column like this:

my_df['some_text_column'] = my_df['some_text_column'].str.replace('~','-')

Then use the first command to confirm that the number of occurences is now 0.

Very similar to above, but with the error message:

bcpandas.constants.BCPandasValueError: Data contains all of the possible delimiter characters (',', '|', '\t'),
cannot use BCP to import it. Replace one of the possible delimiter characters in
your data, or use another method besides bcpandas.

Approach this as is described above for quote characters. If you target the | character for replacement, note that it must be escaped in a regular expression with a backslash. So the relevant commands would be (here the pipe | is replaced with a front slash):

my_df['some_text_column'] = my_df['some_text_column'].str.replace('\|','/')
my_df['some_text_column'].str.contains('\|').sum()

If you get this error message when writing to the database:

Error = [Microsoft][ODBC Driver 17 for SQL Server]Incorrect host-column number found in BCP format-file

Try replacing any space characters in your column names, with a command like my_df.columns = my_df.columns.str.replace(' ','_') (source).

Writing data from pandas DataFrames to a SQL database is very slow using the built-in to_sql method, even with the newly introduced execute_many option. For Microsoft SQL Server, a far far faster method is to use the BCP utility provided by Microsoft. This utility is a command line tool that transfers data to/from the database and flat text files.

This package is a wrapper for seamlessly using the bcp utility from Python using a pandas DataFrame. Despite the IO hits, the fastest option by far is saving the data to a CSV file in the file system and using the bcp utility to transfer the CSV file to SQL Server. Best of all, you don't need to know anything about using BCP at all!

Much credit is due to bcpy for the original idea and for some of the code that was adopted and changed.

The only scope of bcpandas is to read and write between a pandas DataFrame and a Microsoft SQL Server database. That's it. We do not concern ourselves with reading existing flat files to/from SQL - that introduces way to much complexity in trying to parse and decode the various parts of the file, like delimiters, quote characters, and line endings. Instead, to read/write an exiting flat file, just import it via pandas into a DataFrame, and then use bcpandas.

The big benefit of this is that we get to precicely control all the finicky parts of the text file when we write/read it to a local file and then in the BCP utility. This lets us set library-wide defaults (maybe configurable in the future) and work with those.

For now, we are using the non-XML BCP format file type. In the future, XML format files may be added.

• Docker Desktop installed, either of the Linux or Windows runtimes, doesn't matter
• pytest
• hypothesis
• pytest-cov (coverage.py)

We take testing very seriously here. In order to rely on a library like this in production, it MUST be ruthlessly tested, which thankfully it is. Here is a partial list of what has been tested so far. Pull Requests welcome!

• Data types: All ASCII characters 32-127 (using the Hypothesis library, see below). Unicode characters beyond that range have not been tested.
• numpy.NaN, None
• numpy.inf (fails, as expected)
• Empty dataframe (nothing happens, database not modified)
• Duplicate column names (raises error)
• Database columns that are missing from the dataframe, are out of order, or both (passes)
• Extra dataframe columns that aren't in database, when if_exists="append" specified (fails)

• Testing uses pytest.
• To test for all possible data types, we use the hypothesis library, instead of trying to come up with every single case on our own.
• pytest-cov (which uses coverage.py under the hood) is used to measure code coverage. This is then uploaded to codecov.io as part of the CI/CD process (see below).
• In order to spin up a local SQL Server during testing, we use Docker. Specifically, we run one of the images that Microsoft provides that already have SQL Server fully installed, all we have to do is use the image to run a container. Here are the links to the Linux versions and the Windows versions - Express and Developer.
  • When running the tests, we can specify a specific Docker image to use, by invoking the custom command line option called --mssql-docker-image. For example:

    pytest bcpandas/tests --mssql-docker-image mcr.microsoft.com/mssql/server:2019-latest

• Instead of using the subprocess library to control Docker manually, we use the elegant docker-py library which works very nicely. A DockerDB Python class is defined in bcpandas/tests/utils.py and it wraps up all the Docker commands and functionality needed to use SQL Server into one class. This class is used in conftest.py in the core bcpandas tests, and in the benchmarks/ directory for both the benchmarks code as well as the legacy tests for read_sql.

Github Actions is used for CI/CD, although it is still somewhat a work in progress.

Please, all contributions are very welcome!

I will attempt to use the pandas docstring style as detailed here.