[Core] Refactor NodeHead module by optimizing flow insight dependency imports to load only when enabled

[daiping8]

Motivation

NodeHead, as one of the dashboard subprocesses, currently loads a batch of dependencies at process startup that are only needed in specific feature scenarios (especially flow/insight and some autoscaler helper logic). Changing to lazy loading can reduce the base memory footprint of each NodeHead subprocess and shorten import time during startup; in clusters, these savings are amplified by the number of subprocesses.

Changes

• Move flow/insight related heavy dependencies (ray.util.insight / flow_insight / get_resource_usage / PROMPT_TEMPLATE, etc.) from the module top level to inside _emit_node_physical_stats(), importing them only when insight_monitor_address is detected and the feature is enabled.
• Move autoscaler related utility functions (parse_usage, LoadMetricsSummary, get_per_node_breakdown_as_dict) to their actual usage code paths (autoscaler v2 / legacy branches), reducing unnecessary startup-time imports.
• Refactor _to_service_state to a local function to_service_state inside _emit_node_physical_stats(), avoiding top-level symbol coupling for optional dependencies.

Unit Tests

• test_node_head_get_nodes_logical_resources_autoscaler_v2_smoke
  • Force the autoscaler v2 branch (mock is_autoscaler_v2=True)
  • Mock ClusterStatusParser.from_get_cluster_status_reply and parse_usage
  • Verify the return value matches expectations, ensuring no regression of undefined parse_usage

Impact

NodeHead RSS reduced by approximately 10MB

Statistics Script

ps -aux | grep -E 'ray-dashboard' | grep -v grep | awk '{
    rss_mb = $6 / 1024;
    total += rss_mb;
    printf "PID: %-6s RSS: %-8.2f MB CMD: %s\n", $2, rss_mb, $11
} END {
    printf "=====================================\n"
    printf "Total RSS Memory: %-8.2f MB\n", total
}'

[gemini-code-assist]

Code Review

This pull request refactors the NodeHead module to lazily import dependencies for flow insight and autoscaler logic. This is a great optimization to reduce the memory footprint and startup time of the dashboard subprocesses. The changes correctly move imports into the functions where they are used and add checks to only load dependencies when the features are enabled.

I've found one critical issue in the legacy autoscaler code path where a None value could be passed to a function that doesn't expect it, leading to a crash. I've provided a suggestion to fix this.

Overall, the refactoring is well-executed and improves the performance of the dashboard.

[github-actions]

This pull request has been automatically marked as stale because it has not had
any activity for 14 days. It will be closed in another 14 days if no further activity occurs.
Thank you for your contributions.

You can always ask for help on our discussion forum or Ray's public slack channel.

If you'd like to keep this open, just leave any comment, and the stale label will be removed.

[daiping8]

keep alive