Synchronous TM suggestions vs deferred indexing for concurrent Weblate edits

Goal is to keep editor save latency predictable when multiple users edit the same project via Weblate UI or API.

Weblate stores translation units with Django ORM transactions. Background work for commits, translation memory indexing and machine translation suggestions runs via Celery on the same database. With default queues, UI requests can contend for I/O and lock time under parallel load.

Operators face an unresolved trade-off between keeping live translation memory suggestions enabled for immediate editor feedback and deferring or limiting those lookups to reduce request queuing. Weblate does not define a default strategy for live TM search under load.

What configuration determines synchronous versus deferred translation memory suggestion behavior? Is the trade-off documented for PostgreSQL versus SQLite deployments? Which signals indicate that TM work is contributing to latency under concurrent saves?