A Replayed Import Must Not Create Another Relationship

Botie imports a dependency file, but the connection drops before the loader receives a success response. Retrying every row with CREATE can add another copy of relationships already committed. Some rows also refer to assets that have not been imported.

We need stable identities for imported relationships and an explicit route for records that cannot be applied. An import is a reconciliation process, not simply a loop that draws every edge it sees.

Retries, changed reused IDs and missing endpoints require different outcomes.

A relationship requires both endpoints. Import node identities first or resolve them against a known registry. An unmatched endpoint should enter a rejection or deferred-resolution queue with its source row identity.

Silently skipping a MATCH failure can make a loader appear successful while losing relationships. Inventing missing nodes with MERGE can be equally misleading when an ID is misspelled.

Choose whether unresolved endpoints are rejected, deferred or explicitly created; do not let query structure choose by accident.

A stable source relationship ID can identify a logical record across retries. The same ID with the same complete payload is a replay. The same ID with different endpoints, type or properties is a conflict under this lesson’s immutable-import policy.

If the source supports revisions, add a documented version rule rather than treating every changed payload as a harmless repeat. A relationship’s minutes or observation time can matter as much as its endpoints.

Identity equality and payload equality are separate checks.

old={'id':'r1','source':'A','target':'B','type':'ROUTE','minutes':4}
new={**old,'minutes':9}
print(old==new)

False

UNWIND turns a bound list parameter into rows. Matching both endpoint nodes resolves the relationship’s meaning before MERGE uses its stable ID. The example deliberately uses one approved relationship type rather than accepting arbitrary user query structure.

The executor must validate unique IDs and detect changed payload conflicts before this write. It must also compare processed-row counts with the staged batch so unmatched endpoints do not disappear unnoticed.

A successful query response does not prove every source row was imported.

UNWIND $rows AS row
MATCH (a:Asset {id:row.source}), (b:Asset {id:row.target})
MERGE (a)-[r:DEPENDS_ON {import_id:row.id}]->(b)
RETURN row.id AS imported ORDER BY imported

A large import may commit in batches to control memory and transaction size. If batch three fails, batches one and two may already be durable. The loader must resume from a checkpoint that corresponds to committed work, not merely rows read from the file.

Record batch IDs, source versions, accepted counts and rejection counts. Keep checkpoints and sink updates consistent through an appropriate transaction or recovery protocol.

Batching reduces the rollback unit; it does not make the whole source file one atomic transaction.

CSV text needs explicit type conversion and missing-value rules. A blank date is not an arbitrary default instant, and an invalid cost must not enter a weighted path algorithm. Parameter binding protects query structure but does not validate the values’ domain meaning.

Reject or quarantine malformed records with enough non-sensitive context to repair them. Keep credentials, personal data and restricted payloads out of casual loader logs.

Import quality directly changes the answers later traversals can produce.

The exercise returns IDs of new resolvable edges only. The workshop exposes a richer classification before selecting ready rows: repeat, conflict, dangling and invalid records remain inspectable in the intermediate output.

Both mechanisms preserve the original graph. They stage work for a real importer rather than pretending a browser result has committed anything.

Inspect the classification component when an expected ID does not reach the ready list.