Benchmarking

Before you do any benchmarking with APSW or other ways of accessing SQLite, you must understand how and when SQLite does transactions. See transaction control. APSW does not alter SQLite’s behaviour with transactions.

Some access layers try to interpret your SQL and manage transactions behind your back, which may or may not work well with SQLite also doing its own transactions. You should always manage your transactions yourself. For example to insert 1,000 rows wrap it in a single transaction else you will have 1,000 transactions. The best clue that you have one transaction per statement is having a maximum of 60 statements per second. You need two drive rotations to do a transaction - the data has to be committed to the main file and the journal - and 7200 RPM drives do 120 rotations a second. On the other hand if you don’t put in the transaction boundaries yourself and get more than 60 statements a second, then your access mechanism is silently starting transactions for you. This topic also comes up fairly frequently in the SQLite mailing list archives.

speedtest

APSW includes a speed testing script as part of the source distribution. You can use the script to compare SQLite performance across different versions of SQLite, different host systems (hard drives and controllers matter) as well as between pysqlite and APSW. The underlying queries are based on SQLite’s speed test.

$ python speedtest.py --help
Usage: speedtest.py [options]

Options:
  -h, --help           show this help message and exit
  --apsw               Include apsw in testing (False)
  --pysqlite           Include pysqlite in testing (False)
  --correctness        Do a correctness test
  --scale=SCALE        How many statements to execute.  Each unit takes about
                       2 seconds per test on memory only databases. [Default
                       10]
  --database=DATABASE  The database file to use [Default :memory:]
  --tests=TESTS        What tests to run [Default
                       bigstmt,statements,statements_nobindings]
  --iterations=N       How many times to run the tests [Default 4]
  --tests-detail       Print details of what the tests do.  (Does not run the
                       tests)
  --dump-sql=FILENAME  Name of file to dump SQL to.  This is useful for
                       feeding into the SQLite command line shell.
  --sc-size=N          Size of the statement cache. APSW will disable cache
                       with value of zero.  Pysqlite ensures a minimum of 5
                       [Default 100]
  --unicode=UNICODE    Percentage of text that is unicode characters [Default
                       0]
  --data-size=SIZE     Maximum size in characters of data items - keep this
                       number small unless you are on 64 bits and have lots of
                       memory with a small scale - you can easily consume
                       multiple gigabytes [Default same as original TCL
                       speedtest]


$ python speedtest.py --tests-detail
bigstmt:

  Supplies the SQL as a single string consisting of multiple
  statements.  apsw handles this normally via cursor.execute while
  pysqlite requires that cursor.executescript is called.  The string
  will be several kilobytes and with a factor of 50 will be in the
  megabyte range.  This is the kind of query you would run if you were
  restoring a database from a dump.  (Note that pysqlite silently
  ignores returned data which also makes it execute faster).

statements:

  Runs the SQL queries but uses bindings (? parameters). eg::

    for i in range(3):
       cursor.execute("insert into table foo values(?)", (i,))

  This test has many hits of the statement cache.

statements_nobindings:

  Runs the SQL queries but doesn't use bindings. eg::

    cursor.execute("insert into table foo values(0)")
    cursor.execute("insert into table foo values(1)")
    cursor.execute("insert into table foo values(2)")

  This test has no statement cache hits and shows the overhead of
       having a statement cache.

  In theory all the tests above should run in almost identical time
  as well as when using the SQLite command line shell.  This tool
  shows you what happens in practise.