Open Cypher - MERGE doesn't seem to match an existing vertex and tries to create it again

[teojqx]

Environment:

• Ran Docker container on Ubuntu
• Was using version 26.4.2
• All settings are in default

How I set up test schema to reproduce issue (using Studio):

• Create vertex Type "Experiment"
• Add property "pk" to Experiment, set Mandatory and Not Null to true
• Add index on "pk", Index Algorithm is Hash, set Unique to true

Ran this command in Studio using Open Cypher:

MERGE (v:Experiment {pk: "a"})
ON CREATE SET v.status="created"
ON MATCH SET v.status="updated"
RETURN v

First run will create the vertex. Second run of same command will fail with the error Duplicated key [a] found on index 'Experiment[pk]' already assigned to record #80:0 due to the index on "pk" only allowing unique values. I was expecting it to update the "status" property to "updated" but it seems to try and create another vertex.

I've also tested it on "latest" tag which was 26.5.1-SNAPSHOT (build e88e6e4/1778019236199/main) and still see this same issue.

Interestingly, when I tested it on previous 26.3.2 release, there was no issue. The MERGE successfully matches the vertex and updated it accordingly.

[teojqx]

I was curious whether I still see the same issue if there was no index on "pk". I dropped the index on "pk" and tried the same command. This time it was able to match and update the vertex instead of trying to create a new one. So it seems there's no issue with MERGE on non-indexed properties.

I tried another experiment and created the index again but left Unique as false. Now the command will create new vertices since there's no Unique constraint. So MERGE seems to just create new ones if the property to be matched is indexed.

[lvca]

In MergeStep.findAllNodes (engine/src/main/java/com/arcadedb/query/opencypher/executor/steps/MergeStep.java:649), the index lookup path consumed Identifiable items from the cursor and checked instanceof Vertex. Index cursors yield bare RID objects, not Vertex records, so the check always failed and findAllNodes returned empty - making MERGE believe the vertex did not exist and try to create a duplicate, hitting the unique constraint. This affected any unique index (HASH or LSM_TREE) on the matched property; without an index the code falls back to iterateType, which returns actual Vertex records, hence the user's observation that the bug only appeared with an indexed property.

[lvca]

Fix: Resolve the RID to a record via database.lookupByRID(...) and only count vertices, in a new resolveVertex helper.

[lvca]

Fixed!

[teojqx]

Thank you! I've also verified it myself that MERGE is working as I expected with the latest docker image.