performance issues of deep filters

[jexp]

deep filters don't use the indexes but do label scan + filter

i.e. currently we generate ALL(expr IN [(start)-->(next:Next) | next.name = $foo] WHERE expr)

and nested variants of this for each level of nested filters

• we also do NONE/SINGLE/SOME(expr ...)

which we had to generate because:

1. we need the names to express arbitrary predicates and to drill down deeper
2. I didn't think on how to represent ALL except of list of all expression results has to be true

but that leads to indexes not being used

we can change this to: size([(start)-->(next:Next) WHERE next.name = $foo | true]) = 0 where the = 0 depends on the below

• ANY(some): size > 1
• SINGLE: size = 1
• NONE: size = 0
• ALL: size = 0 with inverted condition

this works a bit for single level but for deeper nesting it breaks the planner

one idea is to skip intermediate levels if there are no additional filters

e.g. size([(start)-->(middle)-->(next:Next) WHERE next.name = $foo | true]) = 0

originally we would generate:

size([(start)-->(middle) WHERE size([(middle)-->(next:Next) WHERE next.name = $foo | true]) = 0 | true]) = 0

in 4.0 we can generate existential subqueries to help with that

WHERE exists { MATCH ... WHERE ... } I think will be the syntax.

[jexp]

/cc @DirkMahler

[jexp]

This still uses the index:

profile MATCH (t:User) WITH collect(t) as users
WHERE size([(t)-[:FOLLOWS]->()-[:FOLLOWS]->(u:User) WHERE u.id = "PunkAB" | u]) = 0
RETURN count(*)

[jexp]

This also still uses the index

profile MATCH (t:User) with t limit 10
WHERE size([(t)-[:FOLLOWS]->(b:User) WHERE size([(b)-[:FOLLOWS]->(u:User) WHERE u.id = "PunkAB" | true])=0 | true]) = 0
RETURN count(*)

[jexp]

Problem is with the negation for ALL that kills the index :(

profile MATCH (t:User) with t limit 10
WHERE size([(t)-[:FOLLOWS]->(b:User) WHERE size([(b)-[:FOLLOWS]->(u:User) WHERE NOT u.id = "PunkAB" | true])=0 | true]) = 0
RETURN count(*)

[DirkMahler]

Thanks a lot!