AWS Re:cap 03: Ask Observability Data in Plain Language

AWS Hong Kong Summit 2026 · Developer Lounge Recap · Article 3 of 6

AWS Magazine · Field Report from Victoria Harbour

---

Harbour Lights, High Capital & High Tech: Inside the Developer Lounge

Overlooking the sun-drenched, iconic backdrop of Victoria Harbour, the AWS Hong Kong Summit 2026 Developer Lounge delivered an unmatched synthesis of high-tech innovation and classic Hong Kong luxury. Beneath sleek architectural lighting and surrounded by ambient chill beats, international speakers, chief architects, and founders gathered across plush leather lounges and private espresso bars.

+-----------------------------------------------------------------------------------+
AWS HONG KONG SUMMIT 2026 DEV LOUNGE

[ Harbour View Lounge ] --> [ Hands-On Code Labs ] --> [ VIP Private Meetup ]
  - VC & Founder Deals        - Horizon UI 1.0 Demos      - High-Stakes M&A
  - High-Tech Lux Vibe        - Amazon Bedrock Ops        - HK SRE War Rooms
+-----------------------------------------------------------------------------------+

This wasn't just another tech conference; it was a high-stakes, multi-million-dollar hub of money-in and money-out business collaboration. In quiet corner alcoves, startup founders pitched venture capitalists while cross-border teams hashed out enterprise cloud migration contracts. Yet, at the core of this luxurious atmosphere was an intense, passionate developer culture. Private technical meetups buzzed with debates on site reliability engineering (SRE), agentic workflows, and telemetry architectures.

---

Keynote Spotlight: When the Ops Dashboard Becomes a Conversation

Hong Kong’s SRE culture has long operated at a breakneck, multilingual pace—English in official ticketing systems, Cantonese in high-pressure war rooms, and colorful dashboards everywhere during midnight incidents. At the Summit Developer Lounge, Wu Sheng took the stage to introduce a radically smoother way to manage systems: asking live infrastructure questions in plain language and receiving back trusted, production-grade charts, topologies, and tables.

Summit Headline: "Horizon UI’s new AI Assistant answers questions about your live system with the same charts, topology and tables as the UI — read-only, permission-scoped, and running on a cost-efficient model you bring yourself."

Essential Developer Resources:

● Builder Post: The AI Assistant: Ask Your Observability Data in Plain Language

● Official Documentation: SkyWalking Horizon UI AI Assistant Guide

● Introductory Series: SkyWalking Horizon UI Introduction

This presentation showcased high-tech ops with a luxury product feel: not a clunky, bolted-on chatbot, but an elegant "side drawer of truth" designed to make midnight incident responses structured, calm, and visually clear.

---

The Meet Horizon UI community tour has officially concluded its 17-part series, showcasing every core feature of Apache SkyWalking’s next-generation console, Horizon UI 1.0. The walkthrough highlighted a dynamic sidebar mirroring your entire cloud estate, adaptive dashboards, and 3D topology maps. It also detailed trace, log, and profiling explorers, alongside built-in alarm systems, access control, and highly flexible, config-driven UI customization.

---

From "Meet Horizon UI 17/17" to Horizon UI 1.0

The popular Meet Horizon UI community tour wrapped up its 17/17 series, covering every inch of SkyWalking’s next-generation console:

● The sidebar mirroring your entire cloud estate

● Adaptive dashboards & 3D topology maps

● Trace, log, and profiling explorers

● Alarms, access control, and config-driven UI customization

+-----------------------------------------------------------------------------------+
HORIZON UI EVOLUTION PATH

[ Series 1/17 to 17/17 ]  ------------->  [ Horizon UI 1.0 Release ]
  Full Console Feature Tour                 In-App AI Assistant Launched
  (Dashboards, Topology, Traces)            (Plain Language Live Queries)
+-----------------------------------------------------------------------------------+

Now, Horizon UI 1.0 officially arrives with an in-app AI Assistant that replaces repetitive UI clicking with natural language inquiries.

Plain Language in Action

Instead of clicking through five sub-menus, engineers simply ask:

1. “What’s unhealthy in the system right now?”

2. “Investigate the response time for a service.”

The assistant queries live data from the OAP (Observability Analysis Platform) backend through the exact same path the standard dashboards use, streaming back an ordered narrative built directly from native UI widgets.

Enterprise Security Core Properties:

● Read-Only: Strictly prevents unauthorized write or state changes.

● Permission-Scoped: Directly inherits the logged-in user's role and scopes.

● Off by Default: Remains dormant until explicitly enabled by an administrator and linked to an LLM endpoint.

Lounge Money Angle: Enterprise architects highlighted that bringing your own cost-efficient, tool-calling model eliminates the steep "frontier-model tax" while still delivering executive-ready investigation reports.

---

Hands-On Deep Dive: One Question, a Whole Investigation

When an alert triggers, the AI Assistant triages active alarms, then renders real-time response-time and error-rate figures to explain the anomaly.

+-----------------------------------------------------------------------------------+
AI ASSISTANT INVESTIGATION WORKFLOW

1. Triage Alarms  -->  2. Fetch Live Metrics  -->  3. Render Widgets
   (Active Alerts)        (OAP Backend Query)         (Figure 1, 2, 3)
+-----------------------------------------------------------------------------------+

Visual-First Architecture: Answers You Can See

The assistant follows a strict "show, don't describe" principle. It writes brief explanatory text, draws a real metric figure, interprets the visual trend, and sequentially numbers each block using a running Figure N counter (e.g., "the response-time chart in Figure 1 shows the spike").

Supported Render Formats:

● Line Charts & Single-Value Cards

● Top-N Performance Lists

● Labeled Data Tables & Record Lists

Embedded Dependency Views

When analyzing system relationships, the assistant embeds full, interactive UI components directly into the chat stream:

● One-Hop Topology: Focuses on immediate upstream callers and downstream dependencies.

● Cross-Layer Hierarchy: Projections using the Smartscape fan from service mesh mirrors down to backing infrastructure.

● Deployment Graphs & Instance Maps: Pinpoints specific source-to-destination pairs.

● API-Dependency Chains: Traces complex distributed microservice calls.

Native Signal Explorers

● Traces List: Features a span waterfall view on row clicks (supporting native SkyWalking and Zipkin layers).

● Stored Logs View: Direct access to application log records.

● Browser Errors Stream: Displays frontend error stacks for web applications.

---

Grounded in Live Data: No Metric Hallucinations

The AI Assistant stays strictly grounded by combining three core operational sources:

+-----------------------------------------------------------------------------------+
AI ASSISTANT ENGINE
   --> 1. LIVE DATA
         Reads via OAP query protocols in real-time
   --> 2. LAYER CONFIG
         Metrics catalog, MQE expressions, and scopes
   --> 3. SKYWALKING MODEL
         Scopes, layers, topologies
+-----------------------------------------------------------------------------------+

1. Live System Data: Reads directly via OAP query protocols within the user's specific time window and permissions.

2. Layer Configuration as a Skill: Uses layer/overview templates as a metric catalog. It evaluates verbatim MQE (Metric Query Language) expressions across entities (Service, ServiceInstance, Endpoint) so chat metrics match dashboards perfectly.

3. SkyWalking System Model: Structural understanding of layers, entity scopes, and dependency graphs.

Developer Culture Note: This mirrors the broader Summit 2026 pattern seen in Amazon Bedrock AgentCore Skills—domain expertise lives in structured catalogs and tools, not in hoping an LLM memorizes dynamic system state.

---

Guided Root-Cause Analysis with the Discipline to Stop

When asked to find a root cause, the assistant executes a structured investigation playbook tailored for latency, error rates, saturation, middleware, Kubernetes workloads, or service meshes.

+-----------------------------------------------------------------------------------+
GUIDED ROOT-CAUSE PLAYBOOK FLOW

1. Unhealthy Service
2. Walk Dependencies
3. Isolate Root Service
4. Drill Endpoints / Instances
5. Cross-Layer Hierarchy (mesh / infra)
6. Bottom Out at DB / Queue / Cache Leaf
7. Provide Bounded Answer & kubectl next steps
+-----------------------------------------------------------------------------------+

The Walkthrough Method:

1. Isolates Root Services: Differentiates between a service's own internal errors and symptoms inherited from downstream calls.

2. Drills Deep: Examines the slowest endpoints, instance maps, and exception stacks.

3. Crosses Layers: Follows application dependencies down into backing databases, caches, or message queues.

4. Inspects Infrastructure: Analyzes disk, memory, network edge, and container pressure.

Kubernetes On-Demand Logs & Stopping Discipline

For Kubernetes workloads, the assistant streams container error logs directly into the chat. These logs are never stored permanently, require logs:read permissions, and gracefully inform the user if OAP on-demand log streaming is disabled.

If system data is insufficient to conclusively determine the fault, the assistant provides a bounded, honest summary listing what was proven and what remains unknown—alongside exact kubectl CLI commands for manual verification.

---

Security Architecture & Role-Based Access Control

Safety and compliance are designed directly into the system architecture.

+-----------------------------------------------------------------------------------+
PERMISSION & SAFETY CHECKS

1. User Input
2. Check "ai:read" permission
3. Evaluate Tool Action
4. Re-check granular verb (metrics:read, traces:read, logs:read, ...)
5a. Authorized -> run tool / render figure
5b. Denied     -> show Denied Chip in transcript
+-----------------------------------------------------------------------------------+

Granular Permission Enforcement

Opening the chat drawer requires ai:read. Every individual tool execution subsequently re-verifies its corresponding granular permission:

If a user lacks a specific permission, the assistant renders a clear denied chip in the transcript and continues safely within authorized boundaries.

Profiling: The Single Action Gate

All data collection tools are strictly read-only. Profiling is the only stateful action available, and it requires a two-step confirmation:

1. The assistant presents a decision card explaining why profiling is needed.

2. The user must explicitly approve the action in an interactive popout (and hold profile:enable rights).

Prompt Injection Resilience

The engine treats all incoming telemetry—log strings, service names, trace annotations—as untrusted data to be analyzed, never as executable instructions. A log line containing malicious instructions (e.g., "ignore previous instructions") is simply quoted and analyzed.

---

Bring Your Own Model (BYOM) Setup Guide

Horizon UI 1.0 does not bind you to an expensive proprietary model. Because reasoning logic is encapsulated in tools and playbooks, cost-efficient models operating at temperature = 0 perform exceptionally well.

+-----------------------------------------------------------------------------------+
MODEL INTEGRATION OPTIONS

Option A: Amazon Bedrock Endpoint
  - Enterprise cloud gateway
  - provider: bedrock

Option B: OpenAI-Compatible API
  - Local or hosted gateway
  - provider: openai-compatible
+-----------------------------------------------------------------------------------+

Configuration (horizon.yaml)

You can configure access using horizon.yaml or standard environment variables:

ai:
  enabled: true
  provider: openai-compatible   # openai-compatible or bedrock
  model: "your-cost-effective-model-id"
  baseUrl: "https://your-api-gateway/v1"
  apiKey: "${HORIZON_AI_API_KEY}"   # secret via env only; redacted from logs

Key Deployment Notes:

● Secret Handling: API keys are injected via environment variables, automatically redacted from system logs, and omitted from audit trails.

● UI Workspaces: Access the assistant as a sleek side drawer, expand to a dedicated full page at /ai, or open it in a standalone tab.

● Context & Privacy: Conversation histories are stored locally in the browser, synced across open tabs, and can be cleared at any time from the /ai history tab.

---

Feature Summary: Horizon UI 1.0 AI Assistant

+-----------------------------------------------------------------------------------+
FEATURE SUMMARY MATRIX

Core Model             Vendor-neutral (Amazon Bedrock or OpenAI-compatible)
Key Guardrails         Read-only, least privilege, data/instruction separation
Query Method           Native OAP backend query path (MQE expressions)
UI Capabilities        Live metrics, one-hop topology, waterfall traces, pod logs
Single Active Action   Profiling (requires decision card and user approval)
+-----------------------------------------------------------------------------------+

With Horizon UI 1.0, enterprise observability transitions from manual dashboard diving into clear, interactive conversations—delivering high-tech efficiency directly to modern cloud war rooms.