Disable aggregation expr optimization due to N squared performance

[emilyreff7]

Currently, on instantiation of the Aggregation node, ibis calls into this expr optimization logic, where this line will attempt to simplify a projection by lifting the root. This optimization method is O(n) as it iterates over all the root.selections (here and here), and in practice when called inside the Aggregation constructor, the outcome is O(n^2) as this logic runs for each table column - the result can be very slow expression creation on wide tables when there is a Selection followed by an aggregation.

# create wide table
N = 3000
df = pd.DataFrame({'a': [1, 2, 3]})
cols = {f'col_{i}': [1] * 3 for i in range(N)}
df = df.assign(**cols)
con = Backend().connect({'df': df})
table = con.table('df')

@udf.reduction(input_type=['double'] * N, output_type='double')
def my_reduction_fn(*args, **kwargs):
    return 3.0

# define expr with selection followed by aggregation with all input columns
expr = table.mutate(a=table['a'].cast('float')
all_cols = tuple(expr[col] for col in expr.columns if col != 'a')

%%time
expr = expr.groupby(expr.a).aggregate(result=my_reduction_fn(*all_args))
Wall time: 3.13s

Performance from here is quadratic:

6,000 cols -> Wall time 12.9s  (x2 cols, x2^2 performance)
9,000 cols -> Wall time 28.9 s (x3 cols, x2^3 performance)

For now it is likely better to disable this optimization inside the Aggregation construction until we can improve the performance.

[jreback]

looks fine, can you add a release note & make a tracking issue about this? (e.g. include a ref to the PR summary)