Configuration

Scaler comes with a number of settings that can be tuned for performance. Reasonable defaults are chosen for these that will yield decent performance, but users can tune these settings to get the best performance for their use case. A full list of the available settings can be found by calling the CLI with the -h flag.

Scheduler Settings

For the list of available settings, use the CLI command:

scaler_scheduler -h

Protected Mode

The Scheduler is started in protected mode by default, which means that it can not be shut down by the Client. This is because multiple Clients can connect to a long-running Scheduler, and calling shutdown() may inadverdently kill work that another user is executing. This is also because Scaler encourages strong decoupling between Client, Scheduler, and Workers. To turn off protected mode, start the scheduler with:

scaler_scheduler tcp://127.0.0.1:8516 -p False

Or if using the programmatic API, pass protected=False:

from scaler import SchedulerClusterCombo

cluster = SchedulerClusterCombo(
    address=f"tcp://127.0.0.1:{port}",
    n_workers=2,
    protected=False, # this will turn off protected mode
)

Event Loop

Scaler uses Python’s built-in asyncio event loop by default, however users can choose to use uvloop for the event loop. uvloop is a faster implementation of the event loop and may improve performance of Scaler.

uvloop needs to be installed separately:

pip install uvloop

uvloop can be set when starting the scheduler through the CLI:

scaler_scheduler tcp://127.0.0.1:8516 -e uvloop

from scaler import SchedulerClusterCombo

cluster = SchedulerClusterCombo(
    address=f"tcp://127.0.0.1:{port}",
    n_workers=2,
    event_loop="uvloop"
)

Worker Settings

For the list of available settings, use the CLI command:

scaler_cluster -h

Death Timeout

Workers are spun up with a death_timeout_seconds, which indicates how long the worker will stay alive without being connected to a Scheduler. The default setting is 300 seconds. This is intended for the workers to clean up if the Scheduler crashes.

This can be set using the CLI:

scaler_cluster -n 10 tcp://127.0.0.1:8516 -ds 300

Or through the programmatic API: