Register pivot

[saschahofmann]

Could someone point me to where I could see an example for registering a new BigQuery operator to ibis?

More specifically, I would like to add the new pivot operator.

I imagine it to be called like this

table = con.table('table')
pivot = table.pivot(index='family', column={'type': ['fire', 'water']}, values=['sum', 'strength'])

And I would expect it to produce something like

SELECT * FROM
  (SELECT * FROM `table`)
  PIVOT(SUM(strength) FOR type IN ('fire', 'water', ))

I had a quick look at compiler.py but it isn't clear to me how to achieve this.

[saschahofmann]

I had another look and it seems to me like I would need to overwrite the SelectBuilder and maybe even the UnionBuilder classes. If I wanted to achieve this outside this repo?

[tswast]

The ideal flow is that we add the operator to https://github.com/ibis-project/ibis/blob/master/ibis/expr/operations.py (plus add the pivot method to the table class(es). Then we register the implementation here:

ibis-bigquery/ibis_bigquery/compiler.py

Lines 481 to 484 in abe966b

	# 2.x
	_try_register_op('BitAnd', reduction('BIT_AND'))
	_try_register_op('BitOr', reduction('BIT_OR'))
	_try_register_op('BitXor', reduction('BIT_XOR'))

(assuming this is a new operation in Ibis 2.x).

It's certainly possible to monkeypatch the base ibis classes here to have BigQuery-specific functionality, but I'd like to make sure we stay consistent with the rest of the ibis community.

[saschahofmann]

Hm I think in this case it's very bigquery specific. Most databases don't have a pivot implementation?

[saschahofmann]

I tried to register this for our project. This is what I have so far (probably not quite correct but not the issue right now).

import ibis.expr.rules as rlz
from ibis.expr.operations import Arg, TableNode
from ibis.expr.types import TableExpr
from ibis_bigquery import BigQueryExprTranslator, BigQueryTable
from ibis.expr.schema import HasSchema

compiles = BigQueryExprTranslator.compiles

class Pivot(TableNode, HasSchema):
    table = Arg(TableExpr)
    aggregation = Arg(rlz.scalar(rlz.string))
    input_column = Arg(rlz.noop)
    pivot_columns = Arg(rlz.array_of(rlz.string))

def pivot(table, aggregation, input_column, pivot_columns):
    return Pivot(table, aggregation, input_column, pivot_columns).to_expr()


TableExpr.pivot = pivot

@compiles(Pivot)
def _pivot(t, expr):
    table, aggregation, input_column, pivot_columns = expr.op().args
    return (f"SELECT * FROM ({t.translate(table)})" 
            f"PIVOT({t.translate(aggregation)} FOR {t.translate(input_column)} IN {t.translate(pivot_columns)})")

Unfortunately, when I try to call table.pivot(...).compile() I am getting an error originating in base_sqlalchemy:

.
.
.
~/Projects/gyana/.venv/lib/python3.8/site-packages/ibis/backends/base_sqlalchemy/compiler.py in _collect_elements(self)
    373 
    374         if isinstance(root_op, ops.TableNode):
--> 375             self._collect(source_expr, toplevel=True)
    376             if self.table_set is None:
    377                 raise com.InternalError('no table set')

~/Projects/gyana/.venv/lib/python3.8/site-packages/ibis/backends/base_sqlalchemy/compiler.py in _collect(self, expr, toplevel)
    399             self._collect_Join(expr, toplevel=toplevel)
    400         else:
--> 401             raise NotImplementedError(type(op))
    402 
    403         self.op_memo.add(op)

NotImplementedError: <class '__main__.Pivot'>

I can't figure out what I am doing wrong. It looks to me like it's not registering the compile function . I can call table.pivot(...) without an error but if I call compile it fails. Do I need to do something different when registering something on a TableExpr rather than a StringValue (the only thing I have gotten to work so far).