AIOps and SRE glossary
Definitions for reading the articles and discussing operational decisions. Where a term varies across teams, establish the local meaning before using it as a metric or policy.
Agent loop
A repeated cycle of observation, decision, action, and evaluation. Production implementations need limits, termination conditions, and a handoff when progress or authority is insufficient.
Alert fatigue
Reduced attention to alerts associated with frequent, repetitive, or poorly actionable notifications. It is not defined by a particular false-positive ratio.
AIOps
The application of AI and machine learning to IT operational analysis and workflows. Benefits depend on the use case and evaluation evidence.
Blameless postmortem
An incident review that examines actions in their context and seeks system improvements without assigning blame as its explanatory endpoint. It can still assign ownership for follow-through.
Cardinality
For telemetry, the number of distinct values or label combinations represented. High cardinality can increase storage and query cost; the effect depends on the backend and data model.
Change failure rate
The fraction of production changes that meet the defined failure criterion, such as requiring immediate intervention. Preserve the change population and failure definition when comparing periods.
Chaos engineering
Controlled experimentation that tests assumptions about system behavior under disruptive conditions. Scope, observation, and stop conditions are part of the experiment.
Context window
The bounded amount of tokenized context a model can handle under its interface contract. Input, output, and other token accounting vary by model and API.
Deployment frequency
How often changes are deployed to production over a defined period. Frequency alone does not prove reliability or smaller changes.
DORA metrics
Software delivery performance measures maintained by the DORA research program. Consult the current official guide for its definitions rather than relying on a permanently fixed count of metrics.
Error budget
The allowed amount of bad service compatible with an SLO during its evaluation window. Its units follow the indicator; a request budget is not automatically downtime.
Escalation
Routing a decision or response to the person or team with the needed expertise, capacity, or authority. It need not move upward in a management hierarchy.
Failure injection
Deliberately introducing a specified fault to observe behavior under controlled conditions.
Foundation model
A model trained broadly that can be adapted or applied to multiple downstream tasks. Foundation models are not limited to text or language models.
Incident
An event handled through an organization’s incident process because of actual or threatened service, security, or business impact. Users need not have noticed it first.
LLM
A large language model trained to model and generate language. Its outputs require evaluation against the intended task and can contain unsupported claims.
MLOps
Practices for developing, deploying, evaluating, and operating machine-learning systems, including model and data lifecycle management.
MTTA
Mean time to acknowledge: the average interval from a defined alert or notification event to acknowledgment. State the timestamp boundaries and population.
MTTD
Mean time to detect: the average interval from the defined onset of a problem to detection. Unknown onset times should be disclosed.
MTTF
Mean time to failure: the expected or observed average operating time until failure for the specified population and assumptions. It is distinct from mean time between failures for repairable systems.
MTTR
An ambiguous abbreviation used for repair, recovery, resolution, or response. Define both the expansion and timestamp boundaries before reporting it.
Observability
The ability to infer internal system behavior from available outputs. In software practice, telemetry supports investigation of both anticipated and new questions; it cannot guarantee an explanation.
On-call
Assigned responsibility to respond to specified operational events during a coverage period, which may include ordinary working hours.
P50 / P95 / P99
Percentiles: values at or below which the corresponding fraction of observations falls, subject to the calculation method. Instance percentiles cannot generally be averaged into a fleet percentile.
Playbook
A locally defined operational guide, often covering broader coordination or decision paths. Agree on its distinction from a runbook within the team.
Postmortem
A structured review of an incident’s impact, timeline, contributing conditions, response, and lessons or selected follow-up work.
Prompt engineering
Designing instructions and supplied context to support a model task. Prompt quality can help but does not guarantee correctness or safe execution.
RAG
Retrieval-augmented generation: supplying retrieved material to a generative model as context. Retrieval and citations do not eliminate source errors or unsupported synthesis.
Runbook
A procedure for a defined operational event, with applicability, diagnostic steps, bounded actions, verification, and escalation.
Signal-to-noise ratio
A context-dependent comparison of useful signal with irrelevant or distracting information. For alerts, define actionability and the denominator rather than assuming a universal ratio.
SLA
Service level agreement: an agreement describing service commitments and any associated consequences. Its relationship to internal objectives depends on the agreement.
SLI
Service level indicator: a defined measure of service behavior, with an explicit population, measurement method, and treatment of missing data.
SLO
Service level objective: a target for a defined SLI over an evaluation window.
Telemetry
Data emitted or collected to describe system behavior, including metrics, logs, traces, and other signals.
Token
A unit processed by a model’s tokenizer. It can represent a word fragment, punctuation, or other content; exact accounting depends on the tokenizer and modality.
Token budget
A chosen allowance for token use over a request, workflow, or spending period. Distinguish this policy from the model’s technical context and output limits.
Toil
Operational work characterized by repetition, manual effort, limited enduring value, and growth with service scale. Automation is one possible response; eliminating the need for the work may be better.
Toil budget
A team’s explicit limit or allocation for toil. Separate it from all operational work, and define how the allocation is measured and acted on.
See DORA’s current delivery metrics guide for its maintained definitions, and SRE KPIs: measure outcomes, response, and operational cost for choosing a reliability scorecard.