[
https://issues.apache.org/jira/browse/CALCITE-6939?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17943637#comment-17943637
]
Julian Hyde commented on CALCITE-6939:
--------------------------------------
It seems that the industry has gone with left-to-right. Even though I think
that's a bad idea, we should support it. I'm glad that Mihai added it as a
conformance mode.
I am worried about the validator doing the expansion. We have had many bugs
caused by expansion (e.g. matching columns to GROUP BY columns). I think that
validation could take place without expansion (name resolution, and then type
deduction, by following pointers) and then SQL-to-rel conversion could, again,
follow pointers. Could logic perhaps be included in the namespace rather than
the validator?
> Add support for Lateral Column Alias
> ------------------------------------
>
> Key: CALCITE-6939
> URL: https://issues.apache.org/jira/browse/CALCITE-6939
> Project: Calcite
> Issue Type: Wish
> Components: core
> Affects Versions: 1.39.0
> Reporter: Mihai Budiu
> Priority: Minor
> Labels: pull-request-available
>
> Several SQL dialects allow a SELECT to refer to aliases introduced previously
> in the same select.
> Examples are: Spark, BigQuery, Snowflake. Other names for this feature are
> "column aliasing with immediate reference", or "column alias visibility in
> the same SELECT statement".
> This feature was already proposed and implemented by [CALCITE-5248], but that
> issue and PR contains multiple other unrelated features.
> I am also proposing adding a conformance flag `isSelectAlias` which enables
> this feature. This is strictly a validator feature.
> The semantics has to be clearly defined; we do not want to change the
> semantics of any existing query. My proposal is to allow a name to be
> resolved in the current SELECT statement only if all the other name lookup
> methods have failed.
> I also think it's a good idea to introduce new aliases from left to right, so
> a expression can only refer to aliases introduced to the left of it. This
> removes a few unpleasant corner cases too.
> Another difficult problem is the handling of nondeterministic expressions.
> When you refer to a prior expression by an alias, can/is the expression
> re-evaluated? Unfortunately I believe that the current Calcite IR makes it
> difficult to represent the program in such a way that re-evaluation is
> avoided. (But this is a problem with many other Calcite rewrite rules as
> well.) This feature is still useful when applied to deterministic expressions.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
