pyodbc - very slow bulk insert speed

问题:

With this table:

CREATE TABLE test_insert (
    col1 INT,
    col2 VARCHAR(10),
    col3 DATE
)

the following code takes 40 seconds to run:

import pyodbc

from datetime import date


conn = pyodbc.connect('DRIVER={SQL Server Native Client 10.0};'
    'SERVER=localhost;DATABASE=test;UID=xxx;PWD=yyy')

rows = []
row = [1, 'abc', date.today()]
for i in range(10000):
    rows.append(row)

cursor = conn.cursor()
cursor.executemany('INSERT INTO test_insert VALUES (?, ?, ?)', rows)

conn.commit()

The equivalent code with psycopg2 only takes 3 seconds. I don't think mssql is that much slower than postgresql. Any idea on how to improve the bulk insert speed when using pyodbc?

EDIT: Add some notes following ghoerz's discovery

In pyodbc, the flow of executemany is:

• prepare statement
• loop for each set of parameters
  • bind the set of parameters
  • execute

In ceODBC, the flow of executemany is:

• prepare statement
• bind all parameters
• execute

回答1:

I was having a similar issue with pyODBC inserting into a SQL Server 2008 DB using executemany(). When I ran a profiler trace on the SQL side, pyODBC was creating a connection, preparing the parametrized insert statement, and executing it for one row. Then it would unprepare the statement, and close the connection. It then repeated this process for each row.

I wasn't able to find any solution in pyODBC that didn't do this. I ended up switching to ceODBC for connecting to SQL Server, and it used the parametrized statements correctly.

回答2:

Tried both ceODBC and mxODBC and both were also painfully slow. Ended up going with an adodb connection with help from http://www.ecp.cc/pyado.html. Total run time improved by a factor of 6!

comConn = win32com.client.Dispatch(r'ADODB.Connection')
DSN = 'PROVIDER=Microsoft.Jet.OLEDB.4.0;DATA SOURCE=%s%s' %(dbDIR,dbOut)
comConn.Open(DSN)

rs = win32com.client.Dispatch(r'ADODB.Recordset')
rs.Open('[' + tblName +']', comConn, 1, 3)

for f in values:
    rs.AddNew(fldLST, f)

rs.Update()

回答3:

Trying to insert +2M rows into MSSQL using pyodbc was taking an absurdly long amount of time compared to bulk operations in Postgres (psycopg2) and Oracle (cx_Oracle). I did not have the privileges to use the BULK INSERT operation, but was able to solve the problem with the method below.

Many solutions correctly suggested fast_executemany, however, there are some tricks to using it correctly. First, I noticed that pyodbc was committing after each row when autocommit was set to True in the connect method, therefore this must be set to False. I also observed a non-linear slow down when inserting more than ~20k rows at a time, i.e. inserting 10k rows was subsecond, but 50k was upwards of 20s. I assume that the transaction log is getting quite large and slowing the whole thing down. Therefore, you must chunk your insert and commit after each chunk. I found 5k rows per chunk delivered good performance, but this would obviously depend on many factors (the data, the machine, db config etc...).

import pyodbc

CHUNK_SIZE = 5000

def chunks(l, n):
    """Yield successive n-sized chunks from l."""
    for i in xrange(0, len(l), n): #use xrange in python2, range in python3
        yield l[i:i + n]

mssql_conn = pyodbc.connect(driver='{ODBC Driver 17 for SQL Server}',
                            server='<SERVER,PORT>',
                            timeout=1,
                            port=<PORT>,
                            uid=<UNAME>, 
                            pwd=<PWD>,
                            TDS_Version=7.2,
                            autocommit=False) #IMPORTANT

mssql_cur = mssql_conn.cursor()
mssql_cur.fast_executemany = True #IMPORTANT

params = [tuple(x) for x in df.values]

stmt = "truncate table <THE TABLE>"
mssql_cur.execute(stmt)
mssql_conn.commit()

stmt = """
INSERT INTO <THE TABLE> (field1...fieldn) VALUES (?,...,?)
"""
for chunk in chunks(params, CHUNK_SIZE): #IMPORTANT
    mssql_cur.executemany(stmt, chunk)
    mssql_conn.commit()

回答4:

pyodbc 4.0.19 added a Cursor#fast_executemany option to help address this issue. See this answer for details.

回答5:

I wrote data to text file and then invoked BCP utility. Much much quicker. From about 20 to 30 minutes to a few seconds.

回答6:

I was using pypyODBC w/ python 3.5 and Microsoft SQL Server Management Studio. A particular table ( ~70K rows w/ 40 vars) was taking 112 seconds to INSERT using the .executemany() method with pypyodbc.

With ceODBC it took 4 seconds.