Github user paul-rogers commented on the issue:
https://github.com/apache/drill/pull/1170
Sorry to say, I still disagree with this statement: "This pull request adds
logic to detect and eliminate dangling columns".
There was a prior discussion that `SELECT *` means "return all columns",
not "return only those columns which happen to be in common." We discussed how
removing columns midway through a DAG can produce inconsistent results.
But, let's take this particular case: the Project operator.
What should happen (to be consistent with other parts of Drill), is that
the operator correctly fills in values for the "dangling" F3 so the the output
is (F1, F2, F3).
Note that this becomes VERY ambiguous. Suppose Projection sees the
following from Files A(F1, F2, F3) and B(F1, F2)
* Batch A.1 (with F1, F2, F3)
* Batch B.1 (with F1, F2)
Clearly, the project can remember that F3 was previously seen and fill in
the missing column. (This is exactly what the new projection logic in the new
scan framework does, by the way.) This works, however, only if F3 is nullable.
If not... what (non-null) value can we fill in for F3?
Had we know that F3 would turn up dangling, we could have converted F3 in
the first batch to become nullable, but Drill can't predict the future.
Let's consider the proposal: we drop dangling columns. But, since the
dangling column (F3) appeared in the first batch, we didn't know it is
dangling. Only when we see the second batch (B.1) do we realize that F3 was
dangling and we should have removed it. Again, this algorithm requires
effective time travel.
Now, suppose that the order is reversed:
* Batch B.1 (with F1, F2)
* Batch A.1 (with F1, F2, F3)
Here, we can identify F3 as dangling and could remove it, so the proposal
is sound.
On the other hand, the "fill in F3" trick does not work here because
Project sends B.1 downstream. Later, it notices that A.1 adds a column. Project
can't somehow retroactively add the missing column; all it can do is trigger a
schema change downstream. Again, Drill can't predict the future to know that it
has to fill in F3 in the first B.1 batch.
We've not yet discussed the case in which F2, which exists in both files,
has distinct types (INT in A, say and VARCHAR in B). The dangling column trick
won't work. The same logic as above applies to the type mismatch.
Perhaps we use either the "remove" or "fill in" depending on whether the
column appears in the first batch. So, for the following:
* Batch A.1 (with F1, F2, F3)
* Batch B.1 (with F1, F2)
The result would be (F1, F2, F3)
But if the input was:
* Batch B.1 (with F1, F2)
* Batch A.1 (with F1, F2, F3)
The result wold be (F1, F2)
Since the user has no control over the order that files are read, the
result would be random: half the time the user gets one schema, the other half
the other. It is unlikely that the user will perceive that as a feature.
The general conclusion is that there is no way that Project can "smooth"
the schema in the general case: it would have to predict the future to do so.
Now, let's think about other operators, Sort, say. Suppose we do `SELECT *
FROM foo ORDER BY x`. In this case, there is no project. The Sort operator will
see batches with differing schemas, but must sort/merge them together. The
schemas must match. The Sort tries to do this by using the union type (actually
works, there is a unit test for it somewhere) if columns have conflicting
types. I suspect it does not work if a column is missing from some batches.
(Would need to test.)
And, of course, the situation is worse if the dangling column is the sort
key!
Overall, while it is very appealing to think that dangling columns are a
"bug" for which there is a fix, the reality is not so simple. This is an
inherent ambiguity in the Drill model for which there is no fix that both works
and is consistent with SQL semantics.
What would work? A schema! Suppose we are told that the schema is (F1: INT,
F2: VARCHAR, F3: nullable DOUBLE). Now, when Project (or even the scanner)
notices that F3 is missing, it knows to add in the required column of the
correct type and, voila! no schema change.
Suppose that B defines F2 as INT. We know we want a VARCHAR, and so can do
an implicit conversion. Again, voila! no schema change.
In summary, I completely agree that the scenario described is a problem.
But, I don't believe that removing columns is the fix; instead the only valid
fix is to allow the user to provide a schema.
---
