TechnicalPractitioner3 min read
What is an observability platform made of?
An example self-hostable observability platform, component by component: what each one does, where the data flows, how to watch the platform itself and who owns what.
For: engineers and SREs · architectsPrerequisites: Basic notions of metrics, traces and logs.
An illustrative scenario: an IT department decides to host its own observability rather than pay for an online service. In the steering committee, the CFO asks two simple questions. What exactly are we installing? Who looks after it afterwards? This diagram is there to answer them.
An observability platform reads like a freight network. Applications are the shippers: they produce measurements (metrics), accounts of requests (traces) and journals (logs). Everything arrives at a sorting yard, the OpenTelemetry Collector, which receives, cleans and routes. The data then leaves for specialized warehouses, one per signal type, where people consult it at the counters: dashboards and alerts.
The platform shown here is an example, not a default recommendation. Each card names open source or commercial equivalents. The Collector lets you move from one to another. All components are open source except Phoenix (Elastic License 2.0, source available) and all can run on sovereign infrastructure. Every component is clickable.
Click a component, or reach it with Tab then press Enter, to open its card.
Try this
Section titled “Try this”- Start the animation to follow a signal. The metric, the trace and the log travel together as far as the Collector, in a common format called OTLP. They then split toward three different databases and meet again in Grafana, the dashboard tool. Traces of calls to an AI model (GenAI spans) also go to the tool that evaluates answer quality.
- Open the Collector. Its minimal configuration shows the processing chains, one per signal type, plus a
traces/genaichain that keeps only GenAI spans for the evaluation tool. It also shows the exporters, which send data to each database. This is the only place to change on the day you switch storage backends. - Turn on the “Who owns what” layer. The platform team owns most components, but logs often belong to security and evaluation to the team building the AI. Ask an awkward question: who owns the personal data masking rules in the Collector? And who owns the labels, those tags added to metrics, that blow up the number of series to store?
- Compare network ports. Open Phoenix or Langfuse, then Grafana: two classic clashes appear (port 4317 with the Collector, port 3000 with Grafana). Check them before installing everything on a single host.
What to remember
Section titled “What to remember”Everything goes through the Collector. Applications know a single address. That one place is where you decide where data goes, what gets masked and what is kept.
The platform must also watch itself. Each card gives three health metrics for the component. I recommend gathering them in a pipeline dashboard with alerts attached: silently losing traces is an incident in its own right.
Finally, every component needs an owner’s name. The technical diagram is not enough: without a named person, trade-offs on retention, cardinality or alert thresholds have no decision maker.
Go further: the VictoriaMetrics for LLMs training, the LLM observability training, the GenAI method architecture and the trace explorer to see what this platform carries.
Revised on 2 October 2026: configurations checked against current versions (otlp_grpc and otlp_http exporter names, Tempo 3 retention, GenAI span filter toward the evaluation tool), Phoenix license clarified and alternatives named for each storage component.