LIVE TRACE & SPAN DELIVERY
● STABLE
?
These are simulated live spans (prompts, responses, tool calls, retrieval steps) for 8 monitored LLM applications across 4 simulated enterprise tenants, streaming to this browser at 5 updates per second each. In an Arize AX deployment, Lightstreamer receives these spans from the Kafka trace topic the moment they land, and pushes them instantly to any authorized dashboard, replacing the current micro-batch consumer that waits for up to 100,000 events or a 10-second window. The MERGE channel ensures that under network pressure, a viewer always sees the most recent span state, never a backlog of stale values.
ENTERPRISE TENANT SESSIONS
● STABLE
?
These five sessions represent four simulated enterprise customer dashboard viewers plus one Phoenix self-hosted instance, each on its own network. Lightstreamer's StreamSense transport layer selects the best delivery protocol per session: fast WebSocket for a strong corporate connection, HTTP Streaming for a throttled or proxied one. The network quality indicator here shows the simulated connection strength for each session. A platform operator monitoring all five tenants receives only the latest state from each, with intermediate values intelligently consolidated when bandwidth is limited, exactly the multi-tenant scale question this PoC's Track 2 is built to answer.
EVALUATION SCORES
● STABLE
?
These scores are produced by Arize's evaluation pipeline as new spans arrive from each monitored application. Each score updates at 2 times per second. Lightstreamer delivers each updated score to a viewer's screen within milliseconds of computation, using delta delivery: only the value that changed is transmitted. In a full deployment, Lightstreamer's Kafka Connector bridges the evaluation output topic directly to live dashboards with no custom middleware required. When a hallucination score crosses a threshold, an anomaly alert fires automatically through the DISTINCT channel below.
⚡ LIGHTSTREAMER DELIVERY PROOF
?
The server publishes observability data at a constant rate regardless of any client's network condition. Lightstreamer's MERGE channel tracks the last value delivered to each connected screen and intelligently consolidates intermediate updates when bandwidth is constrained. Each viewer's screen always receives the most recent trace or evaluation state, never a backlog. The conflation ratio shown here is the ratio of server events to client deliveries: 1 to 1 at full bandwidth, rising as network degrades. Anomaly and drift alerts on the DISTINCT channel bypass conflation entirely and always deliver in full.
Data Sources to Lightstreamer
ArizeStreamAdapter publish rate
63
updates/sec (server)
Trace feed (8 apps)5 Hz ea.
Tenant sessions (5)1 Hz ea.
Evaluation scores (8)2 Hz ea.
Platform metrics (2 feeds)1 Hz ea.
Anomaly/drift alertsDISTINCT
MERGE Conflation
?
Conflation ratio: server publish rate divided by client receive rate. At full bandwidth this is 1 to 1. As network degrades, Lightstreamer consolidates intermediate values so the screen always shows the latest trace or evaluation state. Anomaly and drift alerts on the DISTINCT channel bypass conflation and always deliver in full, regardless of bandwidth.
1:1
server events per
client delivery
client delivery
Only the delta is sent.
Unchanged fields transmit 0 bytes.
Unchanged fields transmit 0 bytes.
Lightstreamer to This Screen
This browser's receive rate
63
updates/sec (client)
Trace feed (8 apps)5 Hz ea.
Tenant sessions (5)1 Hz ea.
Evaluation scores (8)2 Hz ea.
Platform metrics (2 feeds)1 Hz ea.
Anomaly/drift alertsALL · DISTINCT
📡 NETWORK SIMULATOR
?
These buttons simulate network conditions an enterprise customer's own dashboard viewer actually encounters: full corporate bandwidth at a desk, a field analyst on a degraded mobile connection, a viewer on a congested or throttled corporate VPN, and a full link drop that simulates a viewer's connection dropping entirely, the same mechanic that validates Phoenix self-hosted stateful reconnect. Lightstreamer adapts automatically to each condition with no configuration required. Use this simulator to observe how the delivery rate and conflation ratio change in the proof panel above, while anomaly and drift alerts continue to arrive regardless of bandwidth.
Effective bandwidth
Unthrottled
Protocol (StreamSense)
WebSocket
MERGE conflation
Off
DISTINCT alerts
Unaffected
Enterprise LAN: WebSocket delivers all observability data at full publish rate. Every span and evaluation score arrives the instant it is published.
⚠ DISTINCT CHANNEL: ANOMALY & DRIFT ALERT DELIVERY
?
Anomaly and drift alerts generated by Arize's evaluation pipeline use Lightstreamer's DISTINCT channel: zero consolidation, every alert delivered exactly once, in sequence, regardless of network conditions. When a viewer's session loses connectivity, alerts queue server-side and flush the instant the connection restores. The published and received counts shown here must always converge: that is the DISTINCT guarantee. This is the channel that ensures a hallucination or prompt-injection alert generated during a VPN drop reaches the on-call reviewer the moment their session reconnects.
Published
?
Total anomaly and drift alerts published by the evaluation pipeline since this session started. This count advances regardless of whether any client is connected.
0
Received
?
Total anomaly and drift alerts received by this browser. During a link loss, this count pauses. On reconnect, queued alerts flush immediately and this count catches up to Published. Published and received must always match.
0
0
alerts queued during last outage
Anomaly and drift alerts bypass all MERGE consolidation. Published and received counts converge after every outage. This is the DISTINCT guarantee: zero missed alerts, zero duplicates, always in order.
📊 PLATFORM COMMAND CENTER
?
An aggregate view of tenant capacity and the Phoenix self-hosted fleet, updating once per second from the same Lightstreamer server that powers the live dashboards. Platform operators, on-call engineers, and account teams each subscribe to the metrics relevant to their role. Lightstreamer delivers each metric the moment it changes, from one server push to as many authorized screens as needed. During a link loss, Lightstreamer holds the last known state server-side and delivers the current values instantly on reconnect.
Platform Ingest
Active Tenants
?
Current active simulated enterprise tenants out of 12 tracked capacity. Updates every second. When active tenant load reaches 90% of tracked capacity, a DISTINCT alert fires to the platform engineering team automatically.
--
of 12 tenants
Consumer Lag
?
Seconds of lag between spans landing on the Kafka trace topic and being available for delivery. This figure updates in real time from the Lightstreamer status monitor. In a full deployment, this is the same lag metric the Week 4 PoC results compile against the current micro-batch baseline.
--
seconds
Capacity
?
Active tenant load as a percentage of tracked platform capacity. At 90% and above, the platform status escalates and a DISTINCT alert notifies the platform engineering team. The bar fills from teal to amber to red as load rises.
--%
utilization
Alerts (1 hr)
?
Total anomaly and drift alerts generated by the evaluation pipeline in the past hour across the platform. Each alert here corresponds to a DISTINCT channel event that was guaranteed delivered to every authorized viewer, including those offline during the alert generation.
--
this hour
Phoenix Self-Hosted Fleet
Instances Active
?
Self-hosted Phoenix instances currently connected. Lightstreamer maintains a persistent connection to each instance, adapting the delivery protocol to the quality of its individual outbound connection, exactly the disconnect-prone environment Track 3 of this PoC is scoped to validate.
--
connected
Avg Fleet Health
?
Average connection health across all active Phoenix self-hosted instances. Lightstreamer's adaptive transport handles the full range from strong to degraded outbound links. Even an instance on a weak connection receives timely evaluation delivery, with conflation ensuring the platform operator always sees the latest state.
--%
quality
⚠ ANOMALY & DRIFT ALERT LOG: DISTINCT CHANNEL
?
Every anomaly or drift alert published by the evaluation pipeline appears here in real time. Alerts shown as QUEUED during a link loss will flip to DELIVERED on reconnect. The sequence number confirms no alert was skipped: DISTINCT guarantees ordered, exactly-once delivery. During a link loss, alerts accumulate server-side and flush in order the instant the connection restores.
Awaiting first anomaly or drift alert...
⚙ LIGHTSTREAMER SESSION LOG
?
Live event stream from the Lightstreamer client library: connection status, subscription acknowledgements, and protocol transitions. WebSocket connected means lowest latency delivery. HTTP Streaming is the fallback for networks that block WebSocket. Smart Polling is the last resort for highly constrained environments. StreamSense selects the right protocol automatically with no configuration required.
Connecting...