Unify and embed multiple filters that operate on the same property #2055
Conversation
andrejtonev
added
Docs - changelog only
andrejtonev
force-pushed
the
range_operator
branch
2 times, most recently
from
1af2133
to
640ce88
Compare
Rewriting the Cypher AST tree to combine multiple filters into a range Insert range filter into range scans
andrejtonev
force-pushed
the
range_operator
branch
from
640ce88
to
3c1c82f
Compare
| @@ -251,6 +252,13 @@ BINARY_OPERATOR_VISIT(SubscriptOperator, "Subscript"); | |||
|
|
|||
| #undef BINARY_OPERATOR_VISIT | |||
|
|
|||
| void ExpressionPrettyPrinter::Visit(RangeOperator &op) { | |||
There was a problem hiding this comment.
How do we reach this code/see its effect?
There was a problem hiding this comment.
This is used when outputting the plan as a JSON.
The JSON is sent as part of the summary when running an EXPLAIN query.
Query (with index):
EXPLAIN MATCH (n:l) WHERE 1 < n.p < 10 RETURN n;
Explain query result:
+---------------------------------------------+
| QUERY PLAN |
+---------------------------------------------+
| " * Produce {n}" |
| " * ScanAllByLabelPropertyRange (n :l {p})" |
| " * Once" |
+---------------------------------------------+
JSON:
"explain": {
"input":{
"input":{
"name":"Once"
},
"label":"l",
"lower_bound":{
"type":"exclusive",
"value":"(ParameterLookup 7)"
},
"name":"ScanAllByLabelPropertyRange",
"output_symbol":"n",
"property":"p",
"upper_bound":{
"type":"exclusive",
"value":"(ParameterLookup 13)"
}
}
}
src/query/plan/preprocess.cpp
Outdated
| @@ -184,7 +186,8 @@ void AddExpansionsToMatching(std::vector<Expansion> &expansions, Matching &match | |||
| } | |||
| } | |||
|
|
|||
| auto SplitExpressionOnAnd(Expression *expression) { | |||
| template <typename F1, typename F2> | |||
| void SplitExpressionOnAnd(Expression *expression, F1 &&if_cb, F2 &&else_cb) { | |||
There was a problem hiding this comment.
Only two usages...inline into both places and refactor them to have simple implementations. The generic callbacks reduce readability.
| @@ -368,6 +369,12 @@ class ReturnBodyContext : public HierarchicalTreeVisitor { | |||
|
|
|||
| #undef VISIT_BINARY_OPERATOR | |||
|
|
|||
| bool PostVisit(RangeOperator &op) override { | |||
| bool res = op.expr1_->Accept(*this); | |||
| res |= op.expr2_->Accept(*this); | |||
There was a problem hiding this comment.
Does 2nd Accept need to be visited?
Can be do a short-cutting ||
There was a problem hiding this comment.
We want to check both branches.
There was a problem hiding this comment.
In the expression a < x < b, a and b do not have to be literals and thus need to be checked.
|
Description
A query like
MATCH(n) WHERE n.p > 10 AND n.p < 20 RETURN n;will generate a tree with 2 independent filters (
n.p > 10andn.p < 20).This becomes a problem when we come to our optimizer.
The optimizer tries to embed filters into range scans. The optimizer will embed one filter, but not the other. It limits the scans to a single filter, so only the upper or lower bound gets embedded.
This PR tries to introduce a new operator RangeOperator.
It should fix this problem and generate a single filter with a lower and upper bound, which gets correctly inserted into a ranged scan.
The operator is not generated by the
CypherMainVisitorbut later as part of the tree rewrite. This is so other optimizations that move filters around can get picked up and potentially combine filters from different parts of the query into a single range.[master < Task] PR
Documentation checklist