title: Web Application description: Servlet container, HTTP traffic, log-level counts, and live Spring AI in-flight operations - direct MeterRegistry reads of operational traffic signals.

Web Application

Web Application - direct MeterRegistry read (no historization through SystemMetricsCollector). The Active LLM ops gauge tracks in-flight ChatClient / Advisor / VectorStore operations as the agent runs.

Purpose - Servlet container, HTTP traffic, logback level counts, and live Spring AI in-flight operations. Different from Host because these metrics are operational traffic signals (rate, in-flight counts, status distribution), not resource consumption.

When to look here

• "Is something blocking HTTP threads?" - HTTP in-flight (server) gauge climbing without proportional throughput.
• "How many concurrent provider calls are in flight right now?" - HTTP in-flight (client) - outbound HTTP to model providers.
• "How many LLM operations are running this second?" - Active LLM ops gauge (ChatClient + Advisor + VectorStore active LongTaskTimers).
• "Are sessions piling up?" - Active sessions + Longest session alive.
• "Did we hit a wave of 4xx / 5xx responses?" - HTTP requests by status chart.
• "Is the WARN/ERROR rate climbing?" - Warn / Error event KPIs + Logback events chart.

Data source

Direct MeterRegistry read (no parallel pipeline - values are live-instant, not historized through SystemMetricsCollector).

Controls

Web Application reads the Observability global refresh interval and ignores the time window - gauges are live, counters are lifetime-cumulative. No tab-specific controls.

KPI cards (thirteen)

Card	Shows	Source
HTTP in-flight (server)	Servlet requests currently being handled	tomcat.threads.busy or LongTaskTimer active count
HTTP in-flight (client)	Outbound HTTP requests in flight (to model providers, MCP servers)	http.client.requests active LongTaskTimer
Active LLM ops	In-flight Spring AI operations	Active LongTaskTimers for chatClient, advisor, vectorStore
Active sessions	Currently active Tomcat sessions	tomcat.sessions.active.current
Longest session alive	Longest-lived active session age	tomcat.sessions.alive.max
Sessions created (lifetime)	Cumulative session create count	tomcat.sessions.created
Sessions expired	Cumulative session expiry count	tomcat.sessions.expired
Sessions rejected	Cumulative session rejection count	tomcat.sessions.rejected
HTTP requests (lifetime)	Cumulative HTTP request count	http.server.requests count
Logback events (lifetime)	Cumulative logback events across all levels	logback.events count
Error events	Cumulative ERROR-level logback events	logback.events{level=ERROR}
Warn events	Cumulative WARN-level logback events	logback.events{level=WARN}
Error/Warn rate	Combined ERROR+WARN rate per minute	Derived from above counters

Charts (four)

Chart	Type	Reading
HTTP requests by status	Horizontal bar (2xx / 3xx / 4xx / 5xx, lifetime)	Sudden 4xx spike → bad request pattern; 5xx → server-side regression
Outbound HTTP latency by host	Horizontal bar (ms by host)	The provider hosts your agent talks to most - useful for diagnosing slow providers
Logback events	Horizontal bar by level (lifetime)	Disproportionate ERROR/WARN → check Logs tab for context
Active LLM operations	Horizontal bar by operation type (ChatClient / Advisor / VectorStore)	Long-running operation types indicate where the agent is currently blocked

Cross-references

• Host - sibling tab for resource consumption (heap / GC / threads / disk)
• Logs - drill into individual log lines when Warn / Error KPI climbs
• Observability Architecture → External export → Metrics - same MeterRegistry is scraped by Prometheus at /actuator/prometheus