Skip to content

4 min read (+1 technical)Video

Labs

The reproducible labs that go with the articles and courses, what is published on the forge and what is still in preparation.

Reading a recipe does not teach you to cook. A configuration is the same: you really understand it by running it, breaking it and fixing it. That is what a lab is for. It is a small, complete environment you start on your own computer, with a scenario to follow.

The site aims for every technical article to link to a lab that reproduces its configuration on a workstation. That is not yet the case everywhere. Here is the actual state.

What is published and what is in preparation

Section titled “What is published and what is in preparation”
LabTopicState
labs/VMLLMVictoriaMetrics as an LLM observability backend, material for the VictoriaMetrics coursepublished on the forge
labs/HPCOBSHPC observability proof of concept, with a training simulatorpublished on the forge
otel/GENAIOTELRAG application instrumented with OpenTelemetry, with an evaluation modulepublished on the forge, see the OpenTelemetry labs
otel/REVERSOTELcontrol plane for a fleet of Collectors through OpAMPpublished on the forge, see the OpenTelemetry labs
otel/OTELAUDITEXPaudit exporter for the Collectorpublished on the forge, see the OpenTelemetry labs
labs/cardinality-cost-labcardinality explosion and its cost, with load profilespublished on the forge, see cardinality explained and its video
labs/slurm-otel-energythe energy used by each job on a GPU cluster, phase by phase, with Slurm and OpenTelemetry; simulated GPUs on a laptop, real NVIDIA GPUs with the same codepublished on the forge, see HPC observability
labs/hpctracefirst version of the previous lab, on a single node: every Slurm job becomes an OpenTelemetry trace, with energy per job and per phase; basis for a semantic conventions proposal for HPCpublished on the forge, see HPC observability
labs/assesmentinterview lab: a connection pool incident to diagnose from traces, metrics and logs, script-driven scenario and interviewer guidepublished on the forge, see the interview
otel/Reversotel-appread-only Grafana app to follow the Collector fleet driven by REVERSOTELpublished on the forge, see the OpenTelemetry labs
ELK/archipeldesign tool: from gathering requirements to the architecture document of an observability, security or search platformpublished on the forge, see Archipel
labs/llm-observabilitykit for LLM observability: the labs (OpenTelemetry, VictoriaMetrics, Tempo, Grafana, Phoenix, Qdrant, Ollama)being fixed, not yet published

Articles that do not have a lab yet say so. A configuration published without having been run end to end carries a box that flags it.

For technical teams: ready-made development environment

Upcoming kits will ship a .devcontainer/ folder; the labs already published on the forge do not have one yet. Once the kit is published, the stack will open in VS Code (Dev Containers), or in GitHub Codespaces if the repository is also published on GitHub.

Archipel: designing the platform before installing it

Section titled “Archipel: designing the platform before installing it”

You do not build a house by pouring the concrete first: you start from what the family needs, then the architect draws the plans. Observability platform projects often skip that step. Teams move quickly from requirements to installation and the design gets lost on the way.

Archipel fills that gap. You describe the need once, without technical vocabulary: the uses, the volumes, how long data is kept, the security and compliance constraints, then the data itself, with the questions you want to ask of it. Archipel derives several scenarios compared side by side, a justified sizing and the architecture document. Every choice points back to the requirement behind it and every rule cites its source.

The name comes from its principle. The need is a neutral base and each technology is an island that translates it into an architecture. Elastic is the only island available today; others are planned, so that several technologies can be compared from the same need.

For technical teams: running Archipel and what is still to be proven
Terminal window
git clone https://forge.erythix.tech/labstraining/ELK/archipel.git
cd archipel
docker compose up -d

The interface then opens on http://127.0.0.1:8765/. The example button loads a fictitious payment institution with every step filled in. Everything runs on the workstation, with no internet access: no project data leaves the machine. Archipel produces the architecture document (overview and detailed view) in PDF, Word or Markdown, along with the deployment code. It also prepares load tests with Rally to check the key assumptions. The features added in October 2026 (sovereignty, centralized deployment, GPU inference) are covered by automated tests but have not yet been used on a real project: the repository says so and lists the facts still to be checked.

The real Grafana dashboard of the HPCOBS lab, filmed live (1 min 36). The data comes from the lab simulator: a 4-GPU training job, then two failures triggered on purpose. GPU utilization stays above 90% while the data loading wait rises and MFU drops; then the gradients explode while the infrastructure looks normal. The lab code is freely available on the forge (public code repository).Music: “Radar Focus”, Blue Saga (Epidemic Sound).

A lab always runs the same way. You get the code, start the environment, trigger an incident, then observe it and fix it. The commands fit in a few lines.

  1. Clone the lab repository from the forge, meaning download a copy of it to your machine.

  2. Start the stack, the set of software that makes up the lab, with docker compose up. No account is required, but you need to download the container images and, for AI labs, the local models, which means several gigabytes on first start.

  3. Follow the scenario in README.md, which has you trigger the incident, observe it and then fix it.

  4. Check the result with the checks listed in README.md, when the lab provides them.