7 min read
Glossary
Observability terms defined simply, with their French equivalent.
Each term is given with its French equivalent, for readers moving between both versions of the site.
ADKAR: change management model published by Prosci, five stages each person goes through: awareness, desire, knowledge, ability, reinforcement.
Alert fatigue (fatigue d’alerte): wear on on-call teams who receive too many alerts, many of them useless. Over time they stop reacting, even when an alert is serious. The Too many alerts: is on-call sustainable? tool measures it.
Blameless post-mortem (post-mortem sans blâme): incident analysis that looks for what made the error possible rather than for a culprit. See the blameless post-mortem.
Burn rate (taux de consommation): how fast an SLO’s error budget is consumed. A burn rate of 1 exhausts the budget exactly at the end of the window.
Cardinality (cardinalité): number of distinct time series a metric produces, that is, unique combinations of its labels. If a metric were a table, this would be its number of rows. A major cost driver of a metrics system.
Chargeback (refacturation interne): actually charging the cost of observability or cloud to each team’s budget.
Collector: program that receives, transforms and forwards traces, metrics and logs, much like a postal sorting center. Usually refers to the OpenTelemetry Collector.
Data contract (contrat de données): written rule that sets, for a signal, its allowed attributes, cardinality budget, retention period and destination. An automatic check verifies it on every code change (continuous integration).
DCGM (Data Center GPU Manager): NVIDIA monitoring tool for GPUs, the processors specialized in heavy computation. Its exporter publishes per-GPU utilization, memory, temperature and power in Prometheus format.
Drift (dérive): progressive degradation of a model in production. Data drift (questions change), concept drift (reality changes) and pipeline drift (a component changes).
eBPF: Linux kernel technology (the kernel is the core of the operating system) that lets you slip small verified programs into it. They attach to hooks (syscalls, functions, network) to observe or filter without changing the applications’ code.
Error budget (budget d’erreur): the share of unreliability an SLO tolerates over a period. A 99.9% SLO over 30 days leaves about 43 minutes of downtime.
Essential entity (entité essentielle): category of organizations under NIS2 with the strictest regime, depending on sector and size; maximum fine of at least €10M or 2% of worldwide turnover.
Exemplar: link from a metric data point to a specific trace, to go from a latency spike to the request that caused it.
Faithfulness (fidélité): share of a response’s claims that are actually backed by the documents retrieved. The core quality metric of a RAG system.
Forge: a platform hosting code repositories (Git), with issue and release tracking, such as GitLab or GitHub. On this site, “the forge” means the Labs & Trainings forge, where the lab code is freely available.
Golden set (jeu de référence): a human-annotated set of questions and answers, used for offline evaluation and to calibrate an LLM judge.
Immutable storage (stockage immuable): storage, also called WORM (write once, read many), where written data can no longer be changed or deleted before the end of its retention period; a necessary condition for evidence, not a sufficient one (see log integrity).
Important entity (entité importante): category of organizations under NIS2 with after-the-fact supervision; maximum fine of at least €7M or 1.4% of worldwide turnover.
Interconnect: very fast network linking the nodes of an HPC cluster, such as InfiniBand.
LLM as a judge (LLM juge): using a language model to score another model’s answers. Fast and reproducible, but subject to known biases.
Log (journal): a dated line describing a one-off event, like an entry in a logbook. It can be written as free text or as fields (structured log).
Metric (métrique): a number recorded at regular intervals, like a meter read every minute, identified by a name and labels.
MIG (Multi-Instance GPU): partitioning of an NVIDIA GPU into instances isolated in compute and memory: each GPU instance has its own memory paths (L2 cache, controllers, bandwidth), according to the NVIDIA documentation. Compute instances created inside the same GPU instance share its memory and engines (MIG concepts).
MTTL (Mean Time To Learn): the average time it takes to understand what is going on, which comes before any fix. The name of this site, echoing MTTR.
MTTR (mean time to recovery): average time to restore service after an incident.
NCCL: NVIDIA library that lets many GPUs exchange data at the same time (AllReduce and broadcast operations). It is used when a model trains on several GPUs at once.
On-call (astreinte): a period during which a person stays reachable outside office hours to step in if a service goes down. In France, it must be compensated. See sustainable on-call.
OTLP (OpenTelemetry Protocol): the transport protocol for traces, metrics, logs and profiles defined by OpenTelemetry.
Payback period (délai de récupération): number of months needed for the cumulative gains of an investment to equal its cost.
Platform engineering (ingénierie de plateforme): building a self-service internal platform, run as a product, that teams use to deploy and run their services.
Profile (profil): regular snapshots of what a program is doing (CPU, memory) to find the part of the code that is expensive. The fourth OpenTelemetry signal, still in development (not stable) as of 2 October 2026 according to the specification status page.
RACI: matrix showing, for each activity, who does the work (R), who is accountable and signs off (A, only one per activity), who is consulted (C) and who is informed (I).
RAG (retrieval augmented generation): first retrieving relevant documents, then passing them to the model so it can use them in its answer, like a student checking their notes before answering.
RDMA (Remote Direct Memory Access): the network adapter reads or writes directly into another machine’s memory, bypassing the operating system kernel, which saves a lot of time; the basis of InfiniBand and RoCE.
SBOM (software bill of materials): the list of software components included in a product, with their versions.
Semantic convention (convention sémantique): standardized names and meanings of OpenTelemetry attributes, such as http.response.status_code or gen_ai.operation.name.
Showback (affichage des coûts): showing each team what it consumes, without charging its budget; the first step towards accountability.
Signal: a type of data a system emits about its own operation. The main ones are metrics (numbers), logs (messages) and traces (the path of a request). Full introduction in observability in 10 minutes.
Silent failure (défaillance silencieuse): a failure with no technical error (hallucination, cost drift, agent loop) that accumulates until it is detected late.
SLI (service level indicator): a measure of service quality as seen by the user, for example the share of requests served under 300 ms.
SLO (service level objective): a target set on an SLI, for example 99.5% of requests under 300 ms over 30 days.
Slurm: the reference HPC scheduler; it queues jobs, allocates resources and accounts for usage per account.
Span: one step in a trace, like a stop in parcel tracking, with a start, an end, attributes and a parent step.
SRE (ingénierie de la fiabilité): site reliability engineering, which treats running a service as an engineering problem; also the person who does it. See profiles.
Straggler (traînard): a GPU or node slower than the others, which makes the whole cluster wait at every synchronization barrier.
Structured interview (entretien structuré): a hiring interview where every candidate gets the same questions, scored with a grid written in advance. See the interview.
Tail sampling (échantillonnage en queue): the decision to keep a trace or not, made once the trace is complete, so knowing how it ended. Keeps every error and only a fraction of normal traffic.
Telemetry (télémétrie): all the signals a system emits, collected continuously and remotely to understand what happens in production. The word comes from remote measurement, as on a satellite or a race car.
Tool poisoning (empoisonnement d’outil): malicious instructions hidden in an MCP tool definition to manipulate the agent that uses it.
Trace: the set of spans describing a request’s path from one service to the next, like tracking a parcel step by step.
TTFT (time to first token): time between sending a request and receiving the first token (the first piece of a word) the model produces. It is the wait the user feels when the answer appears as it is written.
Wilson bound (borne de Wilson): a calculation method that gives a reliable range for a percentage measured on few cases, for example 18 good answers out of 20. Used to compare a quality SLO with its target without false precision.
Revised on 2 October 2026: cardinality described as a major cost driver rather than the first one, MIG definition clarified and sourced from the NVIDIA documentation.