Invalid query generated after series of filters/mutates

[seibs]

When a series of mutates/filters are moved to "WITH" statements, the compiled statement sometimes includes incorrect table references.

Expression

t = ibis.table([('file_date', 'timestamp'), ('PARTITIONTIME', 'date'), ('val', 'int64')])
def test(table):
    table = table[table.PARTITIONTIME < ibis.date('2017-01-01')]
    table = table.mutate(file_date=table.file_date.cast('date'))
    table = table[table.file_date < ibis.date('2017-01-01')]
    table[table]
    table = table.mutate(XYZ=table.val * 2)
    return table
x = test(t)
print(ibis.bigquery.compile(x.join(x.view())[x]))

Compiled

WITH t0 AS (
  SELECT *
  FROM t4
  WHERE `PARTITIONTIME` < '2017-01-01'
),
t1 AS (
  SELECT CAST(`file_date` AS DATE) AS `file_date`, `PARTITIONTIME`, `val`
  FROM t0
),
t2 AS (
  SELECT t1.*, t0.`val` * 2 AS `XYZ`
  FROM t1
  WHERE t1.`file_date` < '2017-01-01'
)
SELECT t2.*
FROM t2
  CROSS JOIN t2 t3

[cpcloud]

This isn't actually a bug, but t4 is appearing because the way we generate names for tables created with ibis.table is the same way we generate aliases for subqueries. If you pass name='foobar' to ibis.table, for example, then you'll get this output:

WITH t0 AS (
  SELECT *
  FROM foobar
  WHERE `PARTITIONTIME` < DATE '2017-01-01'
),
t1 AS (
  SELECT CAST(`file_date` AS DATE) AS `file_date`, `PARTITIONTIME`, `val`
  FROM t0
),
t2 AS (
  SELECT t1.*, t0.`val` * 2 AS `XYZ`
  FROM t1
  WHERE t1.`file_date` < DATE '2017-01-01'
)
SELECT t2.*
FROM t2
  CROSS JOIN t2 t3

This is potentially a duplicate of #1460.

@seibs Is there another issue that manifested this way with non-generated table names?

[seibs]

@cpcloud Sorry, the issue is a little hidden. In the t2 definition, the t0 in the select should be t1.

[cpcloud]

@seibs Finally got around to fixing this. The issue here is that we're fusing across projections without considering the immediate parent table. What happens is that when we're looking for fusion opportunities in a project we look for common expressions in the deepest base table in the tree and assume that we can substitute in the matching columns from that in whatever query we're projecting. This is invalid because there may be other non-fused relations in between. For now, the fix is to only allow fusing with the immediate parent.