Hi, I want to post here in case someone has similar issue with Prefect task.map() and random selection. I have a task which contains random sampling (example below) and the random seed is set for replication purpose. If the task is mapped to multiple config_dict, the result is not consistent - each run will generate different results for the same config and the results are not replicable. If I run the task in a loop (over the multiple config_dict), the result is consistent each time as expected. I am suspecting task.map() starts multi-threads and np.random.seed may not be thread-safe? What's the underlying reason and what's the best practice in this case?

@task
def down_sample(config_dict):
    for k in range(5):
        np.random.seed(seed=k)
        scaling_factors = np.random.choice(
                    f_vector,
                    p=w_vector,
                    size=self.sample_size * self.sample_size,
                    replace=True,)

hi @Manda Li - interesting! submitted / mapped tasks are indeed run in worker threads, so if something is not thread safe that might be causing weirdness do you have an self contained minimal reproduction of this? otherwise I can try to reproduce later