namedtuple row factory for sqlite3

[ncoghlan]

BPO	13299
Nosy	@rhettinger, @ncoghlan, @merwok, @akheron, @serhiy-storchaka, @dlenski, @YoSTEALTH, @erlend-aasland
Files	issue_13299.patch issue_13299.1.patch issue_13299.2.patch sqlite_namedtuplerow.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2011-10-31.00:32:25.001>
labels = ['type-feature']
title = 'namedtuple row factory for sqlite3'
updated_at = <Date 2021-08-05.07:15:53.393>
user = 'https://github.com/ncoghlan'

bugs.python.org fields:

activity = <Date 2021-08-05.07:15:53.393>
actor = 'intellimath'
assignee = 'ghaering'
closed = False
closed_date = None
closer = None
components = []
creation = <Date 2011-10-31.00:32:25.001>
creator = 'ncoghlan'
dependencies = []
files = ['26909', '26945', '26946', '37673']
hgrepos = []
issue_num = 13299
keywords = ['patch']
message_count = 19.0
messages = ['146670', '146738', '147474', '168617', '168748', '168761', '168762', '168763', '218657', '218668', '223704', '223725', '224702', '224707', '233839', '234424', '394912', '395065', '398970']
nosy_count = 11.0
nosy_names = ['rhettinger', 'ghaering', 'ncoghlan', 'eric.araujo', 'petri.lehtinen', 'serhiy.storchaka', 'dlenski', 'Russell.Sim', 'YoSTEALTH', 'erlendaasland', 'intellimath']
pr_nums = []
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue13299'
versions = ['Python 3.5']

[ncoghlan]

Currently, sqlite3 allows rows to be easily returned as ordinary tuples (default) or sqlite3.Row objects (which allow dict-style access).

collections.namedtuple provides a much nicer interface than sqlite3.Row for accessing ordered data which uses valid Python identifiers for field names, and can also tolerate field names which are *not* valid identifiers.

It would be convenient if sqlite3 provided a row factory along the lines of the one posted here:
http://peter-hoffmann.com/2010/python-sqlite-namedtuple-factory.html

(except with smarter caching on the named tuples)

[rhettinger]

+1

[merwok]

collections.namedtuple provides a much nicer interface than sqlite3.Row

Definitely!

[RussellSim]

Hi,

Here is an implementation using lru_cache to prevent regeneration of the named tuple each time.

Cheers,
Russell

[rhettinger]

Caching based on the cursor going to be problematic because a single cursor can be used multiple times with different descriptions:

 c = conn.cursor()
 c.execute('select symbol from stocks')
 print c.description
 c.execute('select price from stocks')
 print c.description # same cursor, different layout, needs a new named tuple

It might make more sense to cache the namedtuple() factory itself:

 sql_namedtuple = lru_cache(maxsize=20)(namedtuple)

Also, the example in the docs is too lengthy and indirect. Cut-out the step for creating an populating the database -- just assume db created in the example at the top of the page:

For example::

 >>> conn.row_factory = sqlite3.NamedTupleRow
 >>> c = conn.cursor()
 >>> for record in c.execute('select * from stocks'):
 print record

 Row(date='2006-01-05', trans='BUY', symbol='RHAT', qty=100.0, price=35.14)
 Row(date='2006-01-05', trans='BUY', symbol='RHAT', qty=100, price=35.14)
 Row(date='2006-03-28', trans='BUY', symbol='IBM', qty=1000, price-45.0)

No need to go into a further lesson on how to use named tuples.

Also, the patch uses star-unpacking: _namedtuple_row(cursor)(*row)

Instead, it should use _make: _namedtuple_row(cursor)._make(row)

(u'2006-04-05', u'BUY', u'MSFT', 1000, 72.0)

[RussellSim]

Raymond, Thanks for the comprehensive feedback! It's fantastic! I have updated the patch with most of you feedback... but there was one part that I couldn't follow entirely. I am now using the _make method but I have had to use star unpacking to allow the method to be cached, lru_cache won't allow a key to be a list because they aren't hash-able.

[ncoghlan]

You should be able to just use "tuple(col[0] for col in cursor.description)" instead of the current list comprehension in order to make the argument hashable.

[RussellSim]

Nick, Thanks for the tip. I have removed the star unpacking.

[glangford]

In abstract, I like the namedtuple interface for sqlite3 as well. One caution is that the approach suggested at

http://peter-hoffmann.com/2010/python-sqlite-namedtuple-factory.html

can have a dramatic impact on performance. For one DB-intensive application, I experienced 20+ seconds run time with the row factory (under 3.4), versus sub second without (identified with cProfile). Many thousands of calls to namedtuple_factory were not good. :)

[glangford]

...if I understand the proposed caching scheme, then repeated executions of the query

SELECT a,b,c FROM table

would result in cache hits, since the column names remain the same. I'm guessing this would resolve the performance problem in the app I saw, but it would be good to verify that performance is broadly similar with/without named tuples.

[BreamoreBoy]

I'd like to see this in 3.5 as I often use sqlite so what needs doing here?

[serhiy-storchaka]

There is significant overhead. Microbenchmark results:

$ ./python -m timeit -s "import sqlite3; con = sqlite3.connect(':memory:')" "con.execute('select 1 as a, 2 as b').fetchall()"
10000 loops, best of 3: 35.8 usec per loop

$ ./python -m timeit -s "import sqlite3; con = sqlite3.connect(':memory:'); con.row_factory = sqlite3.Row" "con.execute('select 1 as a, 2 as b').fetchall()"
10000 loops, best of 3: 37.3 usec per loop

$ ./python -m timeit -s "import sqlite3; con = sqlite3.connect(':memory:'); con.row_factory = sqlite3.NamedTupleRow" "con.execute('select 1 as a, 2 as b').fetchall()"
10000 loops, best of 3: 92.1 usec per loop

It would be easier to add __getattr__ to sqlite3.Row.

[dlenski]

Serhiy,
52 usec/loop doesn't seem like much overhead. This is not 52 usec per row fetched, but just 52 usec per cursor.execute(). An example where >1 row is fetched for each cursor would show this more clearly.

The advantage of namedtuple is that it's a very well-known interface to most Python programmers. Other db-api modules have taken a similar approach; psycopg2 has a dict-like cursor similar to Row, but has added NameTupleCursor in recent versions. (http://initd.org/psycopg/docs/extras.html#namedtuple-cursor)

[serhiy-storchaka]

Yes, above microbenchmarks measure the time of execute() + the time of fetching one row. Here is more precise microbenchmarks.

$ ./python -m timeit -s "import sqlite3; con = sqlite3.connect(':memory:'); con.execute('create table t (a, b)')" -s "for i in range(100): con.execute('insert into t values (1, 2)')" -- "con.execute('select * from t').fetchall()"
1000 loops, best of 3: 624 usec per loop

$ ./python -m timeit -s "import sqlite3; con = sqlite3.connect(':memory:'); con.row_factory = sqlite3.Row; con.execute('create table t (a, b)')" -s "for i in range(100): con.execute('insert into t values (1, 2)')" -- "con.execute('select * from t').fetchall()"
1000 loops, best of 3: 915 usec per loop

$ ./python -m timeit -s "import sqlite3; con = sqlite3.connect(':memory:'); con.row_factory = sqlite3.NamedTupleRow; con.execute('create table t (a, b)')" -s "for i in range(100): con.execute('insert into t values (1, 2)')" -- "con.execute('select * from t').fetchall()"
100 loops, best of 3: 6.21 msec per loop

Here sqlite3.Row is about 1.5 times slower than tuple, but sqlite3.NamedTupleRow is about 7 times slower than sqlite3.Row.

With C implementation of lru_cache() (bpo-14373) the result is much better:

100 loops, best of 3: 3.16 msec per loop

And it will be even more faster (up to 1.7x) when add to the Cursor class a method which returns a tuple of field names.