add "dimension" attribute to pg.ARRAY - implement DDL, and when present optimize bind/result to this

[Mike Bayer]

replaces #1591

[Mike Bayer]

Here's a stack based algorithm to process+copy any N-dimensional structure, given the dimension:

def proc(arr, dim, itemproc):
    dest = [   stack = [(arr, dim, dest)](]
)
    while stack:
        collection, dim, sdest = stack.pop(0)
        if dim == 1:
            for elem in collection:
                sdest.append(itemproc(elem))
        else:
            for elem in collection:
                ssdest = [               stack.append((elem, dim - 1, ssdest))
                sdest.append(ssdest)
    return dest


data1d = [1,2,3](]
)

data3d = [   [
        [1, 2, 3](
),
        [5, 6](4,),
    ],
    [       [7, 8, 9](
),
        [11, 12](10,)
    ]
]

print proc(data1d, 1, lambda x: x+ 1)
print proc(data3d, 3, lambda x: x+ 1)

[Mike Bayer]

p

[Mike Bayer]

I forgot that PG arrays do N-dimensions implicitly. So the attached patch makes "dimensions" optional. Unfortunately we can't do the stack version if we need to convert the collection to a tuple, so we might want to combine all of these approaches, not sure. Needs new tests and docs in any case, the existing tests are all over the place. tests should include dimensions =None to ensure we can still put different numbers of dimensions per individual row.

[Mike Bayer]

let's also make the "collection" configurable - let's make it "collection_cls=<somecls>", astuple does this, then add docs pointing to sqlalchemy.ext.mutable for usage with the ORM, that way no TypeDecorator is needed. probably add a specific example to the mutable docs.

mutable_pg = MutableList.as_mutable(postgresql.ARRAY(Integer, collection_cls=MutableList))

[Mike Bayer]

cfe56e3735d2ba34923c36e9f015253e535ed1bd in 0.8

[Mike Bayer]

Removing milestone: 0.8.0b1 (automated comment)