Hi, I look into the code implementing the rate limit: https://github.com/PrefectHQ/prefect/blob/3bfe04859a2a12c130083a58d4728d043b9c91ed/src/prefect/server/api/concurrency_limits_v2.py

@db_injector
async def bulk_increment_active_slots(
    db: PrefectDBInterface,
    session: AsyncSession,
    concurrency_limit_ids: List[UUID],
    slots: int,
) -> bool:
    active_slots = active_slots_after_decay(db)
    denied_slots = denied_slots_after_decay(db)

    query = (
        sa.update(orm_models.ConcurrencyLimitV2)
        .where(
            sa.and_(
                orm_models.ConcurrencyLimitV2.id.in_(concurrency_limit_ids),
                orm_models.ConcurrencyLimitV2.active == True,  # noqa
                active_slots + slots <= orm_models.ConcurrencyLimitV2.limit,
            )
        )
        .values(
            active_slots=active_slots + slots,
            denied_slots=denied_slots,
        )
    ).execution_options(synchronize_session=False)

    result = await session.execute(query)
    return result.rowcount == len(concurrency_limit_ids)

It looks like the code has a race limit. We could have two transactions coming in parallel from different machines at the same time and as the default transaction level in postgresql is READ COMITTED (https://www.postgresql.org/docs/current/transaction-iso.html) there is no guarantee for serialization. You make the serialization on the client level in a queue but the rate limit can be called globally from multiple levels. Maybe I miss something or maybe this rate limit should work well for 99% of all places, but i would be interested to hear what your motivatins were.