Rewrite LIKE patterns to string matching

[takezoe]

Improve LIKE predicates performance by rewriting to UDFs which don't use a regular expression.

• Add ENDS_WITH and CONTAINS functions
• Rewrite LIKE predicates to STARTS_WITH, ENDS_WITH or CONTAINS if possible

[findepi]

LIKE has special treatment in DomainTranslator.
Currently we do not have to add same treatment for starts_with
because DesugarLikeRewriter is called pretty much at the end
of the optimization process. However, it doesn't have to be always
so. @martint should we have similar support for starts_with in
DomainTranslator?

[takezoe]

I didn't know what DomainTranslator is doing. However, after reading it, I think there is no reason to not apply the same domain translation to STARTS_WITH. I created another pull request for that: #4669

By the way, if Presto has CONTAINS (or MATCHES function which can compare the specified part of string with another string), we will be able to rewrite more patterns of LIKE predicates to not use regular expressions.

[martint]

@martint should we have similar support for starts_with in DomainTranslator?

Yes, that makes sense (for now -- eventually we should be able to get rid of DomainTranslator or drastically reduce the scope of what it does and where it's used).

[takezoe]

Is there a possibility to adopt this kind of optimization? If this optimization has a possibility to be adopted but adding new UDFs is not an ideal solution, I would propose two options:

1. Optimize cases where can be rewritten to STARTS_WITH
2. Introduce a hidden UDF for this optimization

[ubyyj]

hi, what's the reason we haven't applied the optimzation in DesugarLikeRewriter? I think it would help a lot.

[ubyyj]

the rewrite could also happen in ExpressionInterpreter.visitLikePredicate()

[martint]

FYI, as part of #13184, we're trying to remove usage of LikePredicate within the optimizer, as well as all desugaring rules. Those will be part of the initial translation of SQL query AST -> Plan IR.

[colebow]

@martint is it safe to say we can close this PR in favor of the work being done on #13184?

[martint]

It's possible some of the rewrites proposed here are still relevant, but they will need to be reimplemented to match the function calls for the like matching functions instead of the Like expression node, as that's going away in the planner very soon (as part of #13961)

[takezoe]

Sounds like it's better to raise a fresh pull request after #13961. I would close this pull request for now.