As of #3892, the stdout/stderr of all drivers is streamed to all other drivers. An individual driver can opt out by calling ray.init(..., log_to_driver=False).

To reproduce this, do the following.

1. Start Ray.

   ray start --head --redis-port=6379
   

2. Start a driver that prints stuff.

   import ray
   import sys
   import time
   
   ray.init(redis_address='localhost:6379')
   
   @ray.remote
   def f():
       while True:
           print('hi')
           sys.stdout.flush()
           time.sleep(0.1)
   
   f.remote()

3. Start another driver, and see that it receives the stdout from the first driver.

   import ray
   
   ray.init(redis_address='localhost:6379')

Some options to address this.

1. Workers can belong to specific drivers (this means logs will be complicated for cross driver applications, e.g., named actors).
2. Workers can redirect there stdout/stderr on a per-task basis.
3. Workers can print extra annotations in their stdout/stderr which the log monitor can parse and use to choose which drivers to publish to.