The adaption of nested empty list in 2.7.5

[iamalbert]

The adaption of empty list is different in 2.7.4 and 2.7.5

The result of 2.7.5 will cause malformed array literal if it used with =ANY(:arr)

>>> import psycopg2; psycopg2.__version__
'2.7.4 (dt dec pq3 ext lo64)'
>>> from psycopg2.extensions import adapt, register_adapter, AsIs
>>> adapt([]).getquoted()
b"'{}'"
>>> adapt([[]]).getquoted()
b"ARRAY['{}']"
>>> adapt([[[]]]).getquoted()
b"ARRAY[ARRAY['{}']]"

>>> import psycopg2; psycopg2.__version__
'2.7.5 (dt dec pq3 ext lo64)'
>>> from psycopg2.extensions import adapt
>>> adapt([]).getquoted()
b"'{}'"
>>> adapt([[]]).getquoted()
b"'{{}}'"
>>> adapt([[[]]]).getquoted()
b"'{{{}}}'"

[fogzot]

I just added a test for this regression: 1fe9f1a

[dvarrazzo]

For reference, the change was needed to fix #325.

[dvarrazzo]

Observation 1: the previously returned value was wrong:

=# select ARRAY['{}'];
 {"{}"}
=# select ARRAY[ARRAY['{}']];
 {{"{}"}}

This is an array containing the string {}.

Observation 2: postgres just doesn't do arrays of empty arrays:

=# select ARRAY['{}'::text[]];
 {}
=# select ARRAY[ARRAY['{}'::text[]]];
 {}

at this point I have no idea how we are supposed to fix this bug :\

[fogzot]

We can determine if the variable is just a nest of empty arrays and send a single one to PostgreSQL. We lose round-tripping but the value is lost anyway.

[dvarrazzo]

=# select array[array[]]::text[][] = array[]::text[];
 t

This sucks so much... /cringe However it's valid syntax.

[dvarrazzo]

yet another observation the constructor: array[array[]] doesn't work unless it's cast to something:

piro=# select ARRAY[ARRAY[]];
ERROR:  cannot determine type of empty array
LINE 1: select ARRAY[ARRAY[]];
                     ^
HINT:  Explicitly cast to the desired type, for example ARRAY[]::integer[].

this is the same failure of ARRAY[]. Because adapting an empty list is quite desirable we special-case it returning the literal '{}', which I guess it's interpreted as an empty array of unknown in the right context, e.g. in any() (or of the type on the other side of the operator. don't know. don't care).

So, even the test proposed doesn't work:

curs.execute("select null = any(%s)", ([[]], ))
self.assertFalse(curs.fetchone()[0])

this fails because it requires a cast next the placeholder. But it sucks because normally you wouldn't require such a cast: array of non-array don't need it, so you risk a statement working with some values (non-empty lists, empty lists) and failing with some peculiar ones (lists of empty lists). On the other end, if someone has a python list of uniform objects, e.g. ints, I don't expect them to put an empty list into it.

In other words, if the fix-788 branch doesn't show other regressions, I'm ok fixing it this way.