Why is executemany slow in Python MySQLdb?

Question

I am developing a program in Python that accesses a MySQL database using MySQLdb. In certain situations, I have to run an INSERT or REPLACE command on many rows. I am currently doing it like this:

db.execute("REPLACE INTO " + table + " (" + ",".join(cols) + ") VALUES" +
    ",".join(["(" + ",".join(["%s"] * len(cols)) + ")"] * len(data)),
    [row[col] for row in data for col in cols])

It works fine, but it is kind of awkward. I was wondering if I could make it easier to read, and I found out about the executemany command. I changed my code to look like this:

db.executemany("REPLACE INTO " + table + " (" + ",".join(cols) + ") " + 
    "VALUES(" + ",".join(["%s"] * len(cols)) + ")",
    [tuple(row[col] for col in cols) for row in data])

It still worked, but it ran a lot slower. In my tests, for relatively small data sets (about 100-200 rows), it ran about 6 times slower. For big data sets (about 13,000 rows, the biggest I am expecting to handle), it ran about 50 times slower. Why is it doing this?

I would really like to simplify my code, but I don't want the big drop in performance. Does anyone know of any way to make it faster?

I am using Python 2.7 and MySQLdb 1.2.3. I tried tinkering with the setinputsizes function, but that didn't seem to do anything. I looked at the MySQLdb source code and it looks like it shouldn't do anything.

Answer 1

Your first example is a single (large) statement that is generated and then sent to the database.

The second example is a much simpler statement that inserts/replaces a single row but is executed multiple times. Each command is sent to the database separately so you have to pay the turnaround time from client to server and back for every row inserted. I would think that this extra latency introduced between the commands is the main reason for the decreased performance of the second example.

Answer 2

Try lowercasing the word 'values' in your query - this appears to be a bug/regression in MySQL-python 1.2.3.

executemany() matches the VALUES clause with a regular expression and then just clones the list of values for each row of data, so you end up executing exactly the same query as with your first approach.

Unfortunately the regular expression lost its case-insensitive flag in that release (subsequently fixed in trunk r622 but never backported to the 1.2 branch) so it degrades to iterating over the data and firing off a query per row.