beautify default column names

[xudong963]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.
Currently, DataFusion creates column names by https://github.com/apache/arrow-datafusion/blob/7be0e268a69ffecbf06823c98ca572733dddb29e/datafusion/src/physical_plan/planner.rs#L91.
The approach has two existing problems:

1. result in potential bugs, such as A corner bug in union #1818.
2. Makes the printed table unattractive and may be very wide, such as

❯ select COALESCE(null, null, null, null);
+-------------------------------------------------------+
| coalesce(Utf8(NULL),Utf8(NULL),Utf8(NULL),Utf8(NULL)) |
+-------------------------------------------------------+
|                                                       |
+-------------------------------------------------------+
1 row in set. Query took 0.002 seconds.

Describe the solution you'd like

1. Simplify default column names by modifying https://github.com/apache/arrow-datafusion/blob/7be0e268a69ffecbf06823c98ca572733dddb29e/datafusion/src/physical_plan/planner.rs#L91.
2. Directly define a default column name, such as ?column? in postgres:

postgres=# select 1, 2;
 ?column? | ?column?
----------+----------
        1 |        2
(1 row)

[WinkerDu]

I would like to pick this up. Please assign to me as appropriate @xudong963

[xudong963]

Thank you @WinkerDu !

[yjshen]

Please also consider the Apache Hive auto alias as a reference.

TBH, I'm conservative on the need to provide a default name. If something goes wrong, it's acceptable to fail the query, requiring the user to alias explicitly.

[WinkerDu]

@yjshen After trying the following 2 cases in Hive 2.3.7, I got the outputs below:

hive> set hive.cli.print.header=true;
hive> select 1,2;
OK
_c0	_c1
1	2
Time taken: 0.272 seconds, Fetched: 1 row(s)
hive> SELECT coalesce(NULL, 1, NULL);
OK
_c0
1
Time taken: 0.228 seconds, Fetched: 1 row(s)

In my opinion, Hive auto alias convention _c0 seems more reasonable since that indicates column index.
I prefer doing this in Datafusion.
What do you think? @xudong963 @yjshen

[WinkerDu]

Please also consider the Apache Hive auto alias as a reference.

TBH, I'm conservative on the need to provide a default name. If something goes wrong, it's acceptable to fail the query, requiring the user to alias explicitly.

@xudong963 After checking #1818 , like @alamb and @yjshen say, realigning all anonymous Expr seems a bit much.
To solve #1818 , I prefer what @alamb put that adding a LogicalPlan::Projection to Union.
BTW, I'd like to solve that Union corner bug as appropriate, what do you think @xudong963. :D

[xudong963]

@WinkerDu Yeah, you can try to solve #1818 first, we can talk more about #2027

[alamb]

FWIW I think @houqp put a substantial amount of thought into the output name semantics of columns, trying to find a balance between doing what an existing system did and being more useful than ?column?. The details are written in https://arrow.apache.org/datafusion/specification/output-field-name-semantic.html

[gaoxinge]

@alamb Is there any specifications for cast expression as column name?

I raise a pr for cast expression, but it not pass all the tests, and I find my solution is ugly is some cases:

> select cast('1969-01-01T00:00:00.1Z' as timestamp);
"+---------------------------------------------------------------------+",
"| CAST(Utf8(\"1969-01-01T00:00:00.1Z\") AS Timestamp(Nanosecond, None)) |",
"+---------------------------------------------------------------------+",
"| 1969-01-01T00:00:00.100                                             |",
"+---------------------------------------------------------------------+",

[alamb]

@alamb Is there any specifications for cast expression as column name?

I do not know of one, sadly. I think what we have done in the past is to look at other sql implementations (like posgres and spark) for inspiration and followed one of their models

[gaoxinge]

@alamb Thx. I raise an issue #5174 to talk with this case.