neo4j cypher query too slow

Question

The following query, takes between 1.5sec to 9sec, depends on {keywords}

match (pr:Property)
WHERE (pr.name in {keywords})
with pr
MaTCH (pr) <--(it:Item)
MaTCH (it)-->(pr2)<-[:CAT]-(ca)
return distinct pr2 as prop,count(distinct it) as sum , ca.name as rType
limit 10

Each Item is connected to 100 Properties .

sample profile on the server:

neo4j-sh (?)$ profile match (pr:Property)
WHERE (pr.name in ["GREEN","SHORT","PLAIN","SHORT-SLEEVE"])
with pr
MaTCH (pr) <--(it:Item)
MaTCH (it)-->(pr2)<-[:CAT]-(ca)
return distinct pr2 as prop,count(distinct it) as sum , ca.name as rType
limit 40;
+------------------------------------------------------------------------------------------40 rows

ColumnFilter(symKeys=["prop", "rType", "  INTERNAL_AGGREGATE58d28d0e-5727-4850-81ef-7298d63d7be8"], returnItemNames=["prop", "sum", "rType"], _rows=40, _db_hits=0)
Slice(limit="Literal(40)", _rows=40, _db_hits=0)
  EagerAggregation(keys=["Cached(prop of type Node)", "Cached(rType of type Any)"], aggregates=["(  INTERNAL_AGGREGATE58d28d0e-5727-4850-81ef-7298d63d7be8,Distinct(Count(it),it))"], _rows=40, _db_hits=0)
    Extract(symKeys=["it", "ca", "  UNNAMED122", "pr", "pr2", "  UNNAMED130", "  UNNAMED99"], exprKeys=["prop", "rType"], _rows=645685, _db_hits=645685)
      SimplePatternMatcher(g="(it)-['  UNNAMED122']-(pr2),(ca)-['  UNNAMED130']-(pr2)", _rows=645685, _db_hits=0)
        Filter(pred="hasLabel(it:Item(0))", _rows=6258, _db_hits=0)
          SimplePatternMatcher(g="(it)-['  UNNAMED99']-(pr)", _rows=6258, _db_hits=0)
            Filter(pred="any(-_-INNER-_- in Collection(List(Literal(GREEN), Literal(SHORT), Literal(PLAIN), Literal(SHORT-SLEEVE))) where Property(pr,name(1)) == -_-INNER-_-)", _rows=4, _db_hits=1210)
              NodeByLabel(identifier="pr", _db_hits=0, _rows=304, label="Property", identifiers=["pr"], producer="NodeByLabel")

neo4j version : 2.0.1

Heap size : 3.2 GB max (not even close to get to it..)

DataBase disk usage : 270MB

NumOfNodes : 4368

NumOf Relationships : 395693

Computer : AWS EC2 c3.large . But, tried to run it on a 4 times faster computer and the results were the same..

When looking at the JConsole I can see that the heap goes from 50mb to 70mb and then cleaned by GC.

Anyway to make it faster? This performance is way too slow for me...

EDIT: As suggested I tried combining the matches, but it is slower as you can see in the profile:

neo4j-sh (?)$ profile match (pr:Property) WHERE (pr.name in ["GREEN","SHORT","PLAIN","SHORT-SLEEVE"]) with pr MaTCH (pr) <--(it:Item)-->(pr2)<-[:CAT]-(ca) return distinct pr2 as prop,count(distinct it) as sum , ca.name as rType limit 40;

ColumnFilter(symKeys=["prop", "rType", "  INTERNAL_AGGREGATEa6eaa53b-5cf4-4823-9e4d-0d1d66120d51"], returnItemNames=["prop", "sum", "rType"], _rows=40, _db_hits=0)
Slice(limit="Literal(40)", _rows=40, _db_hits=0)
  EagerAggregation(keys=["Cached(prop of type Node)", "Cached(rType of type Any)"], aggregates=["(  INTERNAL_AGGREGATEa6eaa53b-5cf4-4823-9e4d-0d1d66120d51,Distinct(Count(it),it))"], _rows=40, _db_hits=0)
    Extract(symKeys=["  UNNAMED111", "it", "ca", "  UNNAMED119", "pr", "pr2", "  UNNAMED99"], exprKeys=["prop", "rType"], _rows=639427, _db_hits=639427)
      Filter(pred="(hasLabel(it:Item(0)) AND hasLabel(it:Item(0)))", _rows=639427, _db_hits=0)
        SimplePatternMatcher(g="(ca)-['  UNNAMED119']-(pr2),(it)-['  UNNAMED99']-(pr),(it)-['  UNNAMED111']-(pr2)", _rows=639427, _db_hits=0)
          Filter(pred="any(-_-INNER-_- in Collection(List(Literal(GREEN), Literal(SHORT), Literal(PLAIN), Literal(SHORT-SLEEVE))) where Property(pr,name(1)) == -_-INNER-_-)", _rows=4, _db_hits=1210)
            NodeByLabel(identifier="pr", _db_hits=0, _rows=304, label="Property", identifiers=["pr"], producer="NodeByLabel")

Answer 1

First of all, make sure that the name property on the Property label is indexed. As far as I know, indexes aren't used with an IN statement, but this should be resolved in a future version. Performance will be better soon.

CREATE INDEX ON :Property(name)

You can reduce the query as follows:

MATCH (pr:Property)
WHERE (pr.name in {keywords})
MATCH (pr)<--(it:Item)-->(pr2)<-[:CAT]-(ca)
RETURN distinct pr2 as prop,count(distinct it) as sum , ca.name as rType
LIMIT 10

Answer 2

Two you can do as a "workaround", until IN for indexes is fixed:

UNION

split it up in two queries,

first one uses index lookup and a union of all these, like

MATCH (pr:Property {keyword:{keyword1}) return id(pr)
UNION ALL
MATCH (pr:Property {keyword:{keyword2}) return id(pr)
...

etc.

then in the second query do:

MATCH (pr) WHERE ID(pr) IN {ids}
MaTCH (pr) <--(it:Item)
MaTCH (it)-->(pr2)<-[:CAT]-(ca)
return distinct pr2 as prop,count(distinct it) as sum , ca.name as rType
limit 10

Legacy Index

Create a node_auto_index for "keyword" and then use lucene query syntax to do your initial lookup.

START pr=node:node_auto_index('keyword:("GREEN" "SHORT" "PLAIN" "SHORT-SLEEVE")')
MaTCH (pr) <--(it:Item)
MaTCH (it)-->(pr2)<-[:CAT]-(ca)
return distinct pr2 as prop,count(distinct it) as sum , ca.name as rType
limit 10