support path semantic in non-recursive-path

[excaliburwyj]

Description

For non-recursive paths，such as "match (a)-[b]-(c)-[d]-(e) return e;", path semantics do not work.
support has been added:

1. If a path name is not explicitly defined, it will be implicitly created for use in the path semantics filter.
2. An additional path semantic filter plan, such as no_equals of nodes or relationships, will append on top of the hash join or cross process.

Fixes # (issue)

Contributor agreement

• I have read and agree to the Contributor Agreement.

[andyfengHKU]

Hi @excaliburwyj,

Thanks for the contribution. I think the overall idea make sense. Though I'm in middle of migrating recursive join to a different parallel computation framework (#4404). I'll review this in detail during the weekend. Hope this works fine for you.

[andyfengHKU]

Hi @excaliburwyj,

Thanks for the contribution. I think the overall idea make sense. Though I'm in middle of migrating recursive join to a different parallel computation framework (#4404). I'll review this in detail during the weekend. Hope this works fine for you.

Whoops, forgot to review this PR. Let me do it today. Sorry about this.

[andyfengHKU]

Hi @excaliburwyj,

As I'm reviewing this PR, another idea pops into my mind. Instead of first finding a join operator on top of recursive join and then append a filter (which is the logic you implemented in PathSemanticRewriter), I wonder if the following approach will be simpler.

First we add a scalar function hasDuplication in the system. It will be a function that take arbitrary number of nodes, or relations (a relation can be recursive), e.g.

MATCH (a)-[e]->(b)-[]->(c) WHERE NOT hasDuplication(a, b, c)

The above is an equivalent form of

MATCH p = (a)-[e]->(b)-[]->(c)  WHERE is_acyclic(p)

The difference is that it doesn't require a path variable.

Once this is done, for a given MATCH query, we can check if the current semantic is trail or acyclic. If so, and binding stage, we directly add a predicate expression hasDuplication with either nodes or relations depends on the semantic.

In this approach, we can limited change in function module and binder module. And we don't need to worry about planning and optimization.

Let me know what you think.

[excaliburwyj]

Hi @andyfengHKU,
I added the PathSemanticRewriter in the Optimizer mainly for the following reason: From the user's perspective, when I set call recursive_pattern_semantic='TRAIL'/'ACYCLIC'. I would naturally expect the following query to produce different results compared to the 'WALK' mode.
MATCH (a)-[e]->(b)-[]->(c) return c
If must modify the query code by manually adding a filter function, although it can still achieve the path semantics, it could potentially cause confusion. I am curious to know whether you think it is reasonable.