Reordering independent MATCH patterns in ArcadeDB leads to a 33.4x runtime gap

[YaeSakuraQ]

Reordering independent MATCH patterns in ArcadeDB leads to a 33.4x runtime gap

I found two result-equivalent ArcadeDB query formulations whose measured runtime differs substantially on the attached graph. The intended difference between the formulations is a pattern-order variant.

I am reporting the observed runtime/profile difference without assigning it to a specific ArcadeDB component.

ArcadeDB Version: Docker image arcadedata/arcadedb:latest
Operating System: Linux Docker container
Installation Method: Docker image arcadedata/arcadedb:latest
API/Driver: ArcadeDB HTTP OpenCypher endpoint

Steps to reproduce

1. Load the attached graph setup file into an empty disposable ArcadeDB database:

reproducer_graph.cypher.txt

while IFS= read -r stmt; do
  curl -sS -u root:password \
    -H 'Content-Type: application/json' \
    -d "{\"language\":\"opencypher\",\"command\":\"${stmt//\"/\\\"}\"}" \
    http://localhost:2480/api/v1/command/graph
done < reproducer_graph.cypher.txt

2. Verify the graph size:

MATCH (n) RETURN count(n) AS nodes;
-- nodes: 128

MATCH ()-[r]->() RETURN count(r) AS relationships;
-- relationships: 67

3. Run the fast query:

PROFILE
MATCH (n0)
WITH n0
UNWIND [1868710224] AS a0
UNWIND [517248469, -1472958097, 1525636158] AS a1
MATCH (n1 :L1)<-[r0 :T4]-(n2 :L0 :L1)<-[r1 :T1]-(n3), (n4 :L3)
WHERE (((r1.k43) > -2019129302) AND ((r0.id) <> (r1.id)))
WITH a0, (r0.k57) AS a2
ORDER BY a0
RETURN 0 AS a3;

4. Run the slow query:

PROFILE
MATCH (n0)
WITH n0
UNWIND [1868710224] AS a0
UNWIND [517248469, -1472958097, 1525636158] AS a1
MATCH (n4 :L3), (n1 :L1)<-[r0 :T4]-(n2 :L0 :L1)<-[r1 :T1]-(n3)
WHERE (((r1.k43) > -2019129302) AND ((r0.id) <> (r1.id)))
WITH a0, (r0.k57) AS a2
ORDER BY a0
RETURN 0 AS a3;

Expected behavior

The two queries are intended to be result-equivalent. On the same graph, I would expect comparable execution time, or at least not an order-of-magnitude runtime gap, between these two formulations.

Actual behavior

Repeated replay confirmed that the two queries returned the same result rows and reported a large runtime gap.

Metric	Fast query	Slow query
Query form	one result-equivalent pattern-order formulation	reordered result-equivalent pattern-order formulation
Median runtime	11.206 ms	374.813 ms
Runtime ratio	1.000x	33.446x
Runtime delta	n/a	363.607 ms