raydar

A perspective powered, user editable ray dashboard via ray serve.

Features

The raydar module provides an actor which can process collections of ray object references on your behalf, and can serve a perspective dashboard in which to visualize that data.

from raydar import RayTaskTracker
task_tracker = RayTaskTracker()

Passing collections of object references to this actor's process method causes those references to be tracked in an internal polars dataframe, as they finish running.

@ray.remote
def example_remote_function():
    import time
    import random
    time.sleep(1)
    if random.randint(1,100) > 90:
        raise Exception("This task should sometimes fail!")
    return True

refs = [example_remote_function.remote() for _ in range(100)]
task_tracker.process(refs)

This internal dataframe can be accessed via the .get_df() method.

Additionally, setting the enable_perspective_dashboard flag to True in the RayTaskTracker's construction serves a perspective dashboard with live views of your completed references.

task_tracker = RayTaskTracker(enable_perspective_dashboard=True)

Create/Store Custom Views

From the developer console, save your workspace layout locally.

let workspace = document.getElementById('perspective-workspace');

// Save the current layout
workspace.save().then(config => {
    // Convert the configuration object to a JSON string
    let json = JSON.stringify(config);

    // Create a Blob object from the JSON string
    let blob = new Blob([json], {type: "application/json"});

    // Create a download link
    let link = document.createElement('a');
    link.href = URL.createObjectURL(blob);
    link.download = 'workspace.json';

    // Append the link to the document body and click it to start the download
    document.body.appendChild(link);
    link.click();
    document.body.removeChild(link);
});

Then, move this json file to js/src/layouts/default.json.

Expose Ray GCS Information

The data available to you includes much of what Ray's GCS tracks, and also allows for user defined metadata per task.

Specifically, tracked fields include:

• task_id
• user_defined_metadata
• attempt_number
• name
• state
• job_id
• actor_id
• type
• func_or_class_name
• parent_task_id
• node_id
• worker_id
• error_type
• language
• required_resources
• runtime_env_info
• placement_group_id
• events
• profiling_data
• creation_time_ms
• start_time_ms
• end_time_ms
• task_log_info
• error_message

Custom Sources / Update Logic

The proxy server helpd by the RayTaskTracker is exposed via the .proxy_server() property, meaning we can create new tables as follows:

task_tracker = RayTaskTracker(enable_perspective_dashboard=True)
proxy_server = task_tracker.proxy_server()
proxy_server.remote(
    "new",
    "metrics_table",
    {
        "node_id": "str",
        "metric_name": "str",
        "value": "float",
        "timestamp": "datetime",
    },
)

Example: Live Per-Node Training Loss Metrics

If a user were to then update this table with data coming from, for example, a pytorch model training loop with metrics:

def my_model_training_loop()

	for epoch in range(num_epochs):
        # ... my training code here ...

		data = dict(
			node_id=ray.get_runtime_context().get_node_id(),
			metric_name="loss",
			value=loss.item(),
			timestamp=time.time(),
		)
		proxy_server.remote("update", "metrics_table", [data])

Then they can expose a live view at per-node loss metrics across our model training process:

Installation

raydar can be installed via pip or conda, the two primary package managers for the Python ecosystem.

To install raydar via pip, run this command in your terminal:

pip install raydar

To install raydar via conda, run this command in your terminal:

conda install raydar -c conda-forge

License

This software is licensed under the Apache 2.0 license. See the LICENSE file for details.