python sqlite3，我多久要提交一次？

Question

I have a for loop that is making many changes to a database with a sqlite manager class I wrote, but I am unsure about how often I have to commit...

for i in list:
    c.execute('UPDATE table x=y WHERE foo=bar')
    conn.commit()
    c.execute('UPDATE table x=z+y WHERE foo=bar')
    conn.commit()

Basically my question is whether I have to call commit twice there, or if I can just call it once after I have made both changes?

Answer 1

Whether you call conn.commit() once at the end of the procedure of after every single database change depends on several factors.

What concurrent readers see

This is what everybody thinks of at first sight: When a change to the database is committed, it becomes visible for other connections. Unless it is committed, it remains visible only locally for the connection to which the change was done. Because of the limited concurrency features of sqlite , the database can only be read while a transaction is open.

You can investigate what happens by running the following script and investigating its output:

import os
import sqlite3

_DBPATH = "./q6996603.sqlite"

def fresh_db():
    if os.path.isfile(_DBPATH):
        os.remove(_DBPATH)
    with sqlite3.connect(_DBPATH) as conn:
        cur = conn.cursor().executescript("""
            CREATE TABLE "mytable" (
                "id" INTEGER PRIMARY KEY AUTOINCREMENT, -- rowid
                "data" INTEGER
            );
            """)
    print "created %s" % _DBPATH

# functions are syntactic sugar only and use global conn, cur, rowid

def select():
    sql = 'select * from "mytable"'
    rows = cur.execute(sql).fetchall()
    print "   same connection sees", rows
    # simulate another script accessing tha database concurrently
    with sqlite3.connect(_DBPATH) as conn2:
        rows = conn2.cursor().execute(sql).fetchall()
    print "   other connection sees", rows

def count():
    print "counting up"
    cur.execute('update "mytable" set data = data + 1 where "id" = ?', (rowid,))

def commit():
    print "commit"
    conn.commit()

# now the script
fresh_db()
with sqlite3.connect(_DBPATH) as conn:
    print "--- prepare test case"
    sql = 'insert into "mytable"(data) values(17)'
    print sql
    cur = conn.cursor().execute(sql)
    rowid = cur.lastrowid
    print "rowid =", rowid
    commit()
    select()
    print "--- two consecutive w/o commit"
    count()
    select()
    count()
    select()
    commit()
    select()
    print "--- two consecutive with commit"
    count()
    select()
    commit()
    select()
    count()
    select()
    commit()
    select()

Output:

$ python try.py 
created ./q6996603.sqlite
--- prepare test case
insert into "mytable"(data) values(17)
rowid = 1
commit
   same connection sees [(1, 17)]
   other connection sees [(1, 17)]
--- two consecutive w/o commit
counting up
   same connection sees [(1, 18)]
   other connection sees [(1, 17)]
counting up
   same connection sees [(1, 19)]
   other connection sees [(1, 17)]
commit
   same connection sees [(1, 19)]
   other connection sees [(1, 19)]
--- two consecutive with commit
counting up
   same connection sees [(1, 20)]
   other connection sees [(1, 19)]
commit
   same connection sees [(1, 20)]
   other connection sees [(1, 20)]
counting up
   same connection sees [(1, 21)]
   other connection sees [(1, 20)]
commit
   same connection sees [(1, 21)]
   other connection sees [(1, 21)]
$

So it depends whether you can live with the situation that a cuncurrent reader, be it in the same script or in another program, will be off by two at times.

When a large number of changes is to be done, two other aspects enter the scene:

Performance

The performance of database changes dramatically depends on how you do them. It is already noted as a FAQ :

Actually, SQLite will easily do 50,000 or more INSERT statements per second on an average desktop computer. But it will only do a few dozen transactions per second. [...]

It is absolutely helpful to understand the details here, so do not hesitate to follow the link and dive in. Also see this awsome analysis . It's written in C, but the results would be similar would one do the same in Python.

Note: While both resources refer to INSERT , the situation will be very much the same for UPDATE for the same arguments.

Exclusively locking the database

As already mentioned above, an open (uncommitted) transaction will block changes from concurrent connections. So it makes sense to bundle many changes to the database into a single transaction by executing them and the jointly committing the whole bunch of them.

Unfortunately, sometimes, computing the changes may take some time. When concurrent access is an issue you will not want to lock your database for that long. Because it can become rather tricky to collect pending UPDATE and INSERT statements somehow, this will usually leave you with a tradeoff between performance and exclusive locking.