Over a million developers have joined DZone.

Neo4j and Cypher: Profiling ORDER BY LIMIT vs. LIMIT

· Java Zone

Discover how AppDynamics steps in to upgrade your performance game and prevent your enterprise from these top 10 Java performance problems, brought to you in partnership with AppDynamics.

Something I’ve seen people get confused by when writing queries using Neo4j’s cypher query language is the sometimes significant difference in query execution time when using ‘LIMIT’ on its own compared to using it in combination with ‘ORDER BY’.

The confusion is centered around the fact that at first glance it seems like the only thing different between these queries is the sorting of the rows, but there’s actually more to it.

For example, given a small graph with ~ 11,000 nodes…

$ MATCH n RETURN COUNT(n);
 
==> +----------+
==> | COUNT(n) |
==> +----------+
==> | 11442    |
==> +----------+
==> 1 row
==> 1384 ms

…we can return 5 random nodes really quickly with the following query:

$ MATCH n RETURN n LIMIT 5;
 
==> +-------------------------------------------+
==> | n                                         |
==> +-------------------------------------------+
==> | Node[0]{}                                 |
==> | Node[1]{name:"Africa",node_id:"1"}        |
==> | Node[2]{name:"Asia",node_id:"2"}          |
==> | Node[3]{name:"Europe",node_id:"3"}        |
==> | Node[4]{name:"North America",node_id:"4"} |
==> +-------------------------------------------+
==> 5 rows
==> 38 ms

However, if we order those nodes by ID first then it’ll take a bit longer:

$ MATCH n RETURN n ORDER BY ID(n) LIMIT 5;
==> +-------------------------------------------+
==> | n                                         |
==> +-------------------------------------------+
==> | Node[0]{}                                 |
==> | Node[1]{name:"Africa",node_id:"1"}        |
==> | Node[2]{name:"Asia",node_id:"2"}          |
==> | Node[3]{name:"Europe",node_id:"3"}        |
==> | Node[4]{name:"North America",node_id:"4"} |
==> +-------------------------------------------+
==> 5 rows
==> 157 ms

If we have a look at the profile of the second query we can see why this is:

$ PROFILE MATCH n RETURN n ORDER BY ID(n) LIMIT 5;
 
==> ColumnFilter(symKeys=["n", "  UNNAMEDS1215244997"], returnItemNames=["n"], _rows=5, _db_hits=0)
==> Top(orderBy=["SortItem(Cached(  UNNAMEDS1215244997 of type Long),true)"], limit="Literal(5)", _rows=5, _db_hits=0)
==>   Extract(symKeys=["n"], exprKeys=["  UNNAMEDS1215244997"], _rows=11442, _db_hits=0)
==>     AllNodes(identifier="n", _rows=11442, _db_hits=11442)

Top refers to TopPipe and if we look at the internalCreateResults function we can see that the entire collection gets evaluated so that we can sort it correctly on line 63.

if (input.isEmpty)
      Iterator.empty
    else {
      val first = input.next()
      val count = countExpression(first).asInstanceOf[Number].intValue()
 
      val iter = new HeadAndTail(first, input)
      iter.foreach {
        case ctx =>
 
          if (size < count) {
            result += ctx
            size += 1
 
            if (largerThanLast(ctx)) {
              last = Some(ctx)
            }
          } else
            if (!largerThanLast(ctx)) {
              result -= last.get
              result += ctx
              result = result.sortWith((a, b) => compareBy(a, b, sortDescription)(state))
              sorted = true
              last = Some(result.last)
            }
      }
    }

In this case, it means that we load the next node, but in a more complex query, we’d have to evaluate a more complex pattern for every single match.

On the other hand, if we don’t have the ‘ORDER BY’ we use the SlicePipe which does a take on the iterator which lazily retrieves the first 5 results in this instance:

$ PROFILE MATCH n RETURN n LIMIT 5;
 
==> Slice(limit="Literal(5)", _rows=5, _db_hits=0)
==> AllNodes(identifier="n", _rows=5, _db_hits=5)
(skip, limit) match {
      case (Some(x), None) => sourceIter.drop(asInt(x))
      case (None, Some(x)) => sourceIter.take(asInt(x))
 
      case (Some(startAt), Some(count)) =>
        val start = asInt(startAt)
        sourceIter.slice(start, start + asInt(count))
 
      case (None, None) =>
        throw new ThisShouldNotHappenError("Andres Taylor", "A slice pipe that doesn't slice should never exist.")
    }

Now that we’ve drilled into the queries a bit more, we can see that with the ‘ORDER BY LIMIT’ query, many more paths need to be evaluated to make that possible.

A common anti-pattern is to return a result set containing hundreds of thousands of rows and then using ‘ORDER BY LIMIT’ to filter that down to 10.

In that case, there might be a better way of modelling the data so that we have fewer rows coming back from our MATCH clause, and In graph indexes are one such way of doing that.


The Java Zone is brought to you in partnership with AppDynamics. AppDynamics helps you gain the fundamentals behind application performance, and implement best practices so you can proactively analyze and act on performance problems as they arise, and more specifically with your Java applications. Start a Free Trial.

Topics:

Published at DZone with permission of Mark Needham, DZone MVB. See the original article here.

Opinions expressed by DZone contributors are their own.

The best of DZone straight to your inbox.

SEE AN EXAMPLE
Please provide a valid email address.

Thanks for subscribing!

Awesome! Check your inbox to verify your email so you can start receiving the latest in tech news and resources.
Subscribe

{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}