fix(locations): tolerate concurrent creation in bulk_get_or_create

[Maffooch]

Description

A scan import could abort with a database IntegrityError when a second import created the same location row concurrently.

AbstractLocation.bulk_get_or_create looks up existing rows by identity_hash, then bulk-inserts the ones it didn't find. identity_hash is a global singleton with a unique constraint, so when two imports reference the same package/endpoint, one import can commit the row in the window between the other's existence check and its bulk_create. The second import's subtype bulk_create then raised:

duplicate key value violates unique constraint "<subtype>_identity_hash_key"
DETAIL:  Key (identity_hash)=(<hash>) already exists.

That exception propagated out of the location manager's persist() and failed the entire async scan-import task. The previous code deliberately let the transaction roll back on this race, which is exactly what turned a routine collision into a failed import.

Fix

Create the parent Location rows and their subtype rows together inside a savepoint. On a unique-constraint collision the savepoint rolls back — parent Location rows included, so nothing is orphaned — and the enclosing transaction stays usable. The code then re-resolves the rows the concurrent writer committed, carries over their association data, and retries only the rows still missing (bounded retry). A routine collision no longer aborts the import; it resolves to the existing row and continues.

No schema change and no migration: the fix is purely in the create/recover logic, and the existing unique constraint is what makes the recovery correct.

This lives on AbstractLocation, so every location subtype (URL and the Pro Dependency/CodeLocation companions that inherit it) gets the same race tolerance. Pro does not override bulk_get_or_create, so no companion change is required for this path.

Test results

Added a regression test in unittests/test_bulk_locations.py that reproduces the race deterministically: it hides a pre-existing row from the first existence lookup only, so the code attempts a duplicate INSERT and hits the real database unique constraint, then asserts recovery — the raced row resolves to the original (no duplicate created), the genuinely-new row in the same batch is still created, and no orphaned parent Location rows are left behind. The test fails on the previous code (uncaught IntegrityError) and passes with the fix.

Note: this environment has no Docker daemon or Postgres, so the suite was validated through CI rather than locally.

Documentation

No user-facing behavior or documentation change; internal import-robustness fix only.

Checklist

• Bugfix submitted against the bugfix branch.
• Added a regression test to the unit tests.
• No model changes and no migration.
• Ruff compliant.

🤖 Generated with Claude Code

https://claude.ai/code/session_014K826EST31JnjV4xGv6jRf

---

Generated by Claude Code