VictoriaMetrics course glossary

Anomaly score: metric (anomaly_score) produced by vmanomaly, the anomaly detection component of the VictoriaMetrics Enterprise offering, and stored in VictoriaMetrics. A value above 1 indicates an anomaly. It is not a MetricsQL function; with the open source version, a z-score computed in MetricsQL plays a similar role.

Burn rate: speed at which the error budget is consumed. A burn rate of 1 means linear consumption over the SLO window.

Cardinality: number of unique time series (combinations of metric name and label values).

Chunk: document fragment indexed in a vector database for RAG.

Cluster mode (VM): VictoriaMetrics deployment in three components (vminsert, vmstorage, vmselect) for horizontal scaling.

Context pressure: ratio of tokens used to the model’s context window. Above 0.85, truncation becomes a risk.

Counter: metric type that only increases (except on reset). Example: total requests.

DCGM: NVIDIA Data Center GPU Manager. Source of real-time GPU metrics.

Embedding: vector representation of a text, computed by an embedding model.

Error budget: tolerated error margin allowed by the SLO over a given window.

Fan-out: cardinality multiplier resulting from label combinations.

Gauge: metric type that can go up and down. Example: GPU utilization.

Hallucination rate: estimated rate of non-factual responses, measured by evaluators or heuristics.

Histogram: metric type that distributes observations into buckets to estimate quantiles.

Index drift: drift of the retrieval score after reindexing, a sign of an embedding or corpus change.

KV cache: cache of attention keys and values in a Transformer model during inference.

Latency P99: 99th percentile of latency: 99% of requests are faster.

LiteLLM: proxy exposing the APIs of OpenAI, Anthropic, Mistral and others behind a common interface.

MetricsQL: VictoriaMetrics query language, largely compatible with PromQL and enriched with additional functions.

OpenTelemetry (OTel): instrumentation standard for traces, metrics and logs.

OTLP: OpenTelemetry Protocol, over gRPC or HTTP (Protobuf or JSON encoding), used to export telemetry.

Prompt template: prompt skeleton with variables, versioned like code.

RAG: Retrieval Augmented Generation, enriching a prompt with retrieved documents.

Recording rule: rule that pre-computes a derived metric and stores it as a new series.

Remote write: protocol for pushing metrics from Prometheus or vmagent to a remote TSDB.

Retrieval score: similarity score (0 to 1) between a question and the chunks returned by RAG.

Single mode (VM): VictoriaMetrics deployment as a single binary. The vendor documentation recommends it for ingestion rates below about one million data points per second (a recommendation, not a hard limit).

SLA: Service Level Agreement, contractual commitment with an external customer.

SLI: Service Level Indicator, the metric being measured.

SLO: Service Level Objective, internal target on the SLI.

Tenant: logical customer of a multi-tenant platform, isolated by tenant_id.

Token: text unit (often a sub-word) processed by an LLM and billed by providers.

TSDB: Time Series Database. VictoriaMetrics, Prometheus and InfluxDB are TSDBs.

TTFT: Time To First Token, latency before the first token of a streamed response.

vLLM: high-performance open source inference engine for LLMs.

vmagent: VictoriaMetrics component that scrapes targets and sends data with remote write.

vmalert: VictoriaMetrics component that evaluates alerting and recording rules.

vmauth: VictoriaMetrics reverse proxy with authentication and per-user or per-tenant routing.

vmselect: VictoriaMetrics Cluster component that serves queries.

vmstorage: VictoriaMetrics Cluster component that stores data.

vmui: VictoriaMetrics built-in web interface.

VRAM: GPU video memory (HBM, GDDR). Limits the size of the model that can be loaded.

Z-score: number of standard deviations separating a value from the mean of a reference window. Used here to detect retrieval score drift without vmanomaly.

Mark as read