Cloud metrics glossary: what to measure to control performance, availability, and costs
A guide to interpreting latency, errors, CPU, memory, saturation, throughput, and availability, and relating those signals to capacity, scaling, and service commitments.
A platform can remain active and still deliver a poor experience: responses slow down, failures increase, or certain resources approach their limit. For this reason, monitoring the cloud means observing several signals together, not just checking whether a server is on.
This glossary brings together useful indicators for technology and operations teams. Their interpretation depends on the application, the usual load, and the service objectives: an isolated figure does not by itself identify the cause of a problem.
Experience and service metrics
Latency
Latency is the time that elapses between a request and the system's response. It can be measured at different points, such as an API, a database query, or a request handled by the application. If it increases, it is worth investigating which components and dependencies are involved in the operation.
In addition to the average, percentiles such as p95 and p99 are usually observed: they indicate the time below which 95% or 99% of requests, respectively, complete during the measured period. They help detect slow responses that an average may hide. To compare measurements, the interval, unit, and observation point must be kept clear.
Error rate
The error rate expresses what proportion of requests or transactions do not complete successfully. For the indicator to be interpretable, the team must define what counts as an error and what total number of requests is used as the denominator.
An increase may coincide with a slow dependency, an overloaded application, or a recent change. Relating it to latency, the affected services, and the time it started helps narrow the investigation, but it does not by itself prove a cause.
Availability
Availability represents how long a service meets the agreed conditions to be considered operational. It is important to define how it is measured: for example, which tests determine that the service is working and which period is evaluated.
When there is an SLA (service level agreement), availability may form part of its commitments. The indicator and the calculation method must correspond to the agreed conditions; an internal metric does not by itself allow concluding whether the SLA has been met.
Throughput or processed volume
Throughput measures how much work a system handles in a period, for example, requests per second or transactions processed. Interpreting it together with latency and errors allows distinguishing between an increase in activity that the service absorbs well and a load that coincides with degradation.
Editorial illustration · Edition Business
Resource and capacity metrics
CPU usage
CPU usage indicates how much of the available processing resources are being used. High and sustained consumption may signal capacity pressure, although it may also reflect an expected load. Persistently low usage may justify reviewing whether resources are oversized, always taking peaks and the function of the service into account.
The trend and its relationship with the load usually provide more context than a single reading.
Memory usage
Memory usage shows how much memory a process or system consumes relative to the available memory, according to the tool's definition. A growing consumption that does not return to its usual levels may be a signal to investigate; if the pressure continues, it may affect the performance or stability of the service.
It is advisable to analyze its evolution and the application's behavior, rather than assuming that a specific percentage has the same meaning in all environments.
Saturation
Saturation describes how close a resource is to its available capacity or how much pending work it accumulates. It can be observed in CPU, memory, storage, network, database connections, or queues.
It is a pressure signal that may appear before a user perceives an incident. It must be interpreted together with latency, errors, and workload: a busy resource does not necessarily imply a problem if the service continues to handle the load within its objectives.
Scalability and operational efficiency
Autoscaling or automatic scaling
Autoscaling adjusts resources or instances according to configured rules and signals, such as load or CPU usage. In Kubernetes, for example, horizontal scaling can modify the number of replicas of a workload.
Scaling can help absorb demand variations, but it does not automatically fix any bottleneck: an inefficient query, a slow dependency, or limits of another component may continue to condition the service. It is also necessary to observe how resource consumption evolves to understand the operational and economic effect of the changes.
Load balancing
Load balancing distributes requests among several servers or instances. It can reduce the concentration of work on a single node and contribute to availability, provided that the components are configured to handle the load and that the balancer correctly detects health problems.
Bottlenecks and dependencies
A bottleneck appears when a component limits the performance of the whole. It may be in compute, network, storage, a database, or an external dependency. Service and resource metrics, observed in a correlated way, help locate where the operation deteriorates.
The proximity of the infrastructure to users or to the systems it communicates with, as well as the use of cache and CDN, can also influence latency. These measures do not replace measurement: their effect must be checked in the actual path of requests.
A practical framework for interpreting the signals
To provide context for each indicator, teams can document:
What is measured: service, component, operation, and unit.
In what period: aggregation interval and observation window.
What behavior is expected: usual load, peaks, and service objectives.
With which other signals it is compared: for example, latency, errors, throughput, and saturation.
What operational action it enables: investigating a dependency, reviewing capacity, or checking a recent change.
This common definition helps teams interpret alerts consistently and distinguish a one-off variation from a trend. It also allows relating performance to efficiency: the same signals that reveal a lack of capacity can help detect underutilized resources, without an isolated reading being enough to decide changes.
A step-by-step procedure for SMEs to identify applicable legislation, distinguish harmonized standards, and manage conformity assessment, technical documentation, and CE marking before marketing in the European Union.
Clear definitions of the AfCFTA, Regional Economic Communities, and rules of origin to help SMEs interpret the regulatory framework and plan their entry into the African market.
Unifying technical terminology between purchasing, logistics, and quality departments is key to reducing operational errors and improving traceability. This glossary synthesizes fundamental concepts such as the cold chain, cross-docking, ABC inventory management, and ESG criteria, facilitating more precise and efficient communication in the agri-food sector.