Zero/inaccurate relationships count from apoc.meta.schema

[Shady-Hegazy]

For the below bug report, we exemplify using the Geography DB from the CypherBench collection of datasets.Schema file.

Expected Behavior

apoc.meta.schema() and apoc.meta.schema({sample: -1}) should return accurate relationship count values per node label, where:

• Outgoing counts are non-zero when relationships exist.
• Source-side (out) and target-side (in) counts for the same directed relationship are equal.
• Per-label-pair counts reflect only relationships between that specific label pair, not global totals.

Actual Behavior

Both commands produce inaccurate count values. Four problems observed:

1. Outgoing counts consistently reported as 0 for high-cardinality source labels (in Geography domain e.g., Mountain, River, Lake) even when the target-side incoming count is non-zero. It seems to occur exclusively for outgoing relationships.
2. Consequently: source/target counts are contradictory for the same relationship (e.g., flowsThrough: River out → 0, Country in → 82,155).
3. Per-label-pair counts equal global totals — the same global count is reported for multiple distinct target labels simultaneously (e.g., both Lake.flowsInto and Ocean.flowsInto report 75,139).
4. Sampled mode (without 'sample:-1')severely undercounts (e.g., flowsInto: sampled 1,032 vs full-scan 75,139).

How to Reproduce the Problem

1. Load and start a dataset that has directed relationships.
2. Run CALL apoc.meta.schema({sample: -1})
3. Compare reported counts for 'OUT' relationships against actual counts in the DB. Plus, check the duplication (reuse of global total counts) for relationships that are defined for different node pairs.
4. Run CALL apoc.meta.schema() and compare sampled vs full-scan counts

Simple Dataset (where it's possibile)

Geography dataset in https://github.com/megagonlabs/cypherbench. Neo4j docker images are available.

Steps (Mandatory)

1. Load and start a dataset that has directed relationships.
2. Run CALL apoc.meta.schema({sample: -1})
3. Compare reported counts for 'OUT' relationships against actual counts in the DB. Plus, check the duplication (reuse of global total counts) for relationships that are defined for different node pairs.
4. Run CALL apoc.meta.schema() and compare sampled vs full-scan counts

Specifications

Version: 2026.04.0
APOC version: 2026.04.0-core
OS: macOS Tahoe 26.3.1
Interface: Neo4j desktop 2
Protocol: neo4j://

• OS:
• Neo4j:
• Neo4j-Apoc:

[gem-neo4j]

Thanks for writing in! I believe I have found and fixed the bug. The fix should be included in the 2026.06 release :)