wehub-resource-sync
189 lines
6.5 KiB
JSON
Executable File
189 lines
6.5 KiB
JSON
Executable File
{
|
|
"agent":{
|
|
"metrics_collection_interval":60,
|
|
"run_as_user":"root"
|
|
},
|
|
"logs":{
|
|
"metrics_collected": {
|
|
"prometheus": {
|
|
"log_group_name": "{cluster_name}-ray-prometheus",
|
|
"prometheus_config_path": "/opt/aws/amazon-cloudwatch-agent/etc/prometheus.yml",
|
|
"emf_processor": {
|
|
"metric_declaration_dedup": true,
|
|
"metric_namespace": "{cluster_name}-ray-prometheus",
|
|
"metric_unit":{
|
|
"python_gc_collections_total": "Count",
|
|
"python_gc_objects": "Count",
|
|
"python_gc_objects_uncollectable_total": "Count",
|
|
"python_gc_objects_collected_total": "Count",
|
|
"ray_cluster_active_nodes": "Count",
|
|
"ray_cluster_pending_nodes": "Count",
|
|
"ray_node_cpu_count": "Count",
|
|
"ray_node_cpu_utilization": "Percent",
|
|
"ray_node_disk_free": "Bytes",
|
|
"ray_node_disk_usage": "Bytes",
|
|
"ray_node_disk_utilization_percentage": "Percent",
|
|
"ray_node_mem_available": "Bytes",
|
|
"ray_node_mem_total": "Bytes",
|
|
"ray_node_mem_used": "Bytes",
|
|
"ray_node_mem_total_host": "Bytes",
|
|
"ray_node_mem_used_host": "Bytes",
|
|
"ray_node_network_receive_speed": "Bytes",
|
|
"ray_node_network_received": "Bytes",
|
|
"ray_node_network_send_speed": "Bytes",
|
|
"ray_node_network_sent": "Bytes",
|
|
"ray_avg_num_executed_tasks": "Count",
|
|
"ray_avg_num_scheduled_tasks": "Count",
|
|
"ray_avg_num_spilled_back_tasks": "Count",
|
|
"ray_object_manager_num_pull_requests": "Count",
|
|
"ray_object_store_available_memory": "Bytes",
|
|
"ray_object_store_used_memory": "Bytes",
|
|
"ray_object_store_fallback_memory":"Bytes",
|
|
"ray_object_store_num_local_objects": "Count",
|
|
"ray_object_directory_subscriptions": "Count",
|
|
"ray_object_directory_added_locations": "Count",
|
|
"ray_object_directory_removed_locations": "Count",
|
|
"ray_object_directory_lookups": "Count",
|
|
"ray_object_directory_updates": "Count",
|
|
"ray_pending_actors": "Count",
|
|
"ray_pending_placement_groups": "Count",
|
|
"ray_raylet_cpu": "Count",
|
|
"ray_raylet_mem": "Bytes",
|
|
"ray_raylet_mem_uss": "Bytes",
|
|
"ray_workers_mem": "Bytes",
|
|
"ray_workers_mem_uss": "Bytes",
|
|
"ray_internal_num_spilled_tasks": "Count",
|
|
"ray_internal_num_infeasible_tasks": "Count",
|
|
"ray_internal_num_processes_started": "Count",
|
|
"ray_internal_num_received_tasks": "Count",
|
|
"ray_internal_num_dispatched_tasks": "Count",
|
|
"process_max_fds": "Count",
|
|
"process_open_fds": "Count",
|
|
"process_resident_memory_bytes": "Bytes",
|
|
"process_virtual_memory_bytes": "Bytes",
|
|
"process_start_time_seconds": "Seconds",
|
|
"process_cpu_seconds_total": "Seconds",
|
|
"autoscaler_config_validation_exceptions": "Count",
|
|
"autoscaler_node_launch_exceptions": "Count",
|
|
"autoscaler_pending_nodes": "Count",
|
|
"autoscaler_reset_exceptions": "Count",
|
|
"autoscaler_running_workers": "Count",
|
|
"autoscaler_started_nodes": "Count",
|
|
"autoscaler_stopped_nodes": "Count",
|
|
"autoscaler_update_loop_exceptions": "Count",
|
|
"autoscaler_worker_create_node_time": "Seconds",
|
|
"autoscaler_worker_update_time": "Seconds",
|
|
"autoscaler_updating_nodes": "Count",
|
|
"autoscaler_successful_updates": "Count",
|
|
"autoscaler_failed_updates": "Count",
|
|
"autoscaler_failed_create_nodes": "Count",
|
|
"autoscaler_recovering_nodes": "Count",
|
|
"autoscaler_successful_recoveries": "Count",
|
|
"autoscaler_failed_recoveries": "Count"
|
|
},
|
|
"metric_declaration": [
|
|
{
|
|
"source_labels": [
|
|
"job"
|
|
],
|
|
"label_matcher": "ray",
|
|
"dimensions": [
|
|
[
|
|
"instance"
|
|
]
|
|
],
|
|
"metric_selectors": [
|
|
""
|
|
]
|
|
}
|
|
]
|
|
}
|
|
}
|
|
},
|
|
"logs_collected":{
|
|
"files":{
|
|
"collect_list":[
|
|
{
|
|
"file_path":"/tmp/ray/session_*/logs/**.out",
|
|
"log_group_name":"{cluster_name}-ray_logs_out",
|
|
"log_stream_name":"{instance_id}"
|
|
},
|
|
{
|
|
"file_path":"/tmp/ray/session_*/logs/**.err",
|
|
"log_group_name":"{cluster_name}-ray_logs_err",
|
|
"log_stream_name":"{instance_id}"
|
|
}
|
|
]
|
|
}
|
|
}
|
|
},
|
|
"metrics": {
|
|
"namespace": "{cluster_name}-ray-CWAgent",
|
|
"aggregation_dimensions": [
|
|
[
|
|
"InstanceId"
|
|
]
|
|
],
|
|
"append_dimensions": {
|
|
"AutoScalingGroupName": "${aws:AutoScalingGroupName}",
|
|
"InstanceId": "${aws:InstanceId}"
|
|
},
|
|
"metrics_collected": {
|
|
"collectd": {
|
|
"metrics_aggregation_interval": 60
|
|
},
|
|
"cpu": {
|
|
"measurement": [
|
|
"usage_active",
|
|
"usage_system",
|
|
"usage_user",
|
|
"usage_idle",
|
|
"time_active",
|
|
"time_system",
|
|
"time_user",
|
|
"time_idle"
|
|
],
|
|
"resources": [
|
|
"*"
|
|
]
|
|
},
|
|
"processes": {
|
|
"measurement": [
|
|
"processes_running",
|
|
"processes_sleeping",
|
|
"processes_zombies",
|
|
"processes_dead",
|
|
"processes_total"
|
|
],
|
|
"metrics_collection_interval": 60,
|
|
"resources": [
|
|
"*"
|
|
]
|
|
},
|
|
"disk": {
|
|
"measurement": [
|
|
"disk_used_percent"
|
|
],
|
|
"metrics_collection_interval": 60,
|
|
"resources": [
|
|
"/"
|
|
]
|
|
},
|
|
"mem": {
|
|
"measurement": [
|
|
"mem_used_percent"
|
|
],
|
|
"metrics_collection_interval": 60,
|
|
"resources": [
|
|
"*"
|
|
]
|
|
},
|
|
"statsd": {
|
|
"metrics_aggregation_interval": 60,
|
|
"metrics_collection_interval": 10,
|
|
"service_address": ":8125"
|
|
}
|
|
}
|
|
}
|
|
}
|