Docs: align bottleneck troubleshooting guidance with new job/fragment CPU-memory-busy metrics

[kwannoel]

Context

Internal Slack thread (2026-03-18, #product discussion) requested docs updates after Grafana/dashboard changes:

• Storage alerts/metrics were refined to reduce noise.
• Job/fragment-level CPU, memory, and busy-rate metrics are now available.
• Dashboard tables now show bottleneck percentages by job.
• Team feedback: Jarvis/docs still often tell users to rely on backpressure first, which may no longer match current troubleshooting expectations.

Related references:

• https://github.com/risingwavelabs/risingwave-docs/blob/main/performance/troubleshoot-high-latency.mdx
• https://github.com/risingwavelabs/risingwave-docs/blob/main/performance/metrics.mdx

Problem

Current user-facing troubleshooting guidance appears inconsistent with the latest PROD Grafana workflow. Users may still be guided to use backpressure-centric debugging as the primary path, while the dashboard now exposes direct job/fragment-level CPU/memory/busy-rate and bottleneck ranking tables intended for faster bottleneck localization.

This can cause:

• slower triage due to outdated diagnostic sequence,
• confusion between dashboard capabilities and docs/Jarvis recommendations,
• inconsistent support guidance across engineers.

Suggested Fix

Update troubleshooting + metrics docs to reflect a modern, explicit workflow:

1. Start from job/fragment bottleneck tables and job/fragment CPU/memory/busy-rate panels to identify likely hotspots.
2. Use backpressure metrics as secondary corroboration and graph-propagation analysis (not the only default entry point).
3. Add a short “when to use what” decision table:
   • resource saturation suspected -> CPU/memory/busy-rate + bottleneck tables
   • downstream blockage propagation suspected -> backpressure/path tracing
4. Ensure Jarvis-facing guidance snippets in docs are aligned with this flow so generated support responses follow the same order.
5. Add/update screenshots for the new memory/fragment/job-level panels and bottleneck tables.