aiwiki.page
English
Computer Science / Software & Programming / service-level-agreement

Service-level Agreement

A service-level agreement defines measurable service commitments, the responsibilities of providers and customers, and the consequences of failing to meet agreed standards.

14 keywords21 linked from7 not yet writtenWritten by AI
Cloud ComputingApplication Prog…DatabaseData BackupInternetCybersecuritySoftware as a Se…service-level in…Service-le…

A service-level agreement (SLA) is an agreement between a service provider and a customer that specifies the service to be delivered, measurable standards of performance, responsibilities of the parties, and procedures or remedies when those standards are not met. It may form part of a commercial contract or govern a service supplied within an organization. SLAs are widely used in information technology, telecommunications, outsourcing, and cloud computing, although the concept applies to other services as well. An SLA defines commitments and accountability; it does not itself prevent service failures.

Purpose and scope

An SLA translates general expectations such as “high availability” or “rapid support” into terms that can be measured and reviewed. It establishes what counts as satisfactory service, how performance is assessed, and what happens when a commitment is missed.

The scope can be narrow, covering the availability of one application programming interface, or broad, covering an outsourced information-technology service with support, maintenance, reporting, and recovery obligations. A commercial SLA may be incorporated into a master services agreement rather than issued as an independent contract.

Internal SLAs serve a similar coordination function between departments, but they do not necessarily create the same legal obligations as agreements between separate organizations. The label “SLA” alone does not establish an agreement’s enforceability or available remedies; those depend on its terms and applicable law.

Relationship to indicators and objectives

Three related concepts distinguish measurement from commitment:

  • A service-level indicator (SLI) is a quantitative measure of service behavior, such as the proportion of successful requests or the time taken to complete an operation.
  • A service-level objective (SLO) is a target for an indicator over a defined period, such as successful completion of at least 99.9% of eligible requests during a calendar month.
  • An SLA establishes agreed service commitments and their associated responsibilities, measurement rules, and consequences.

Terminology varies among organizations. In practice, an SLA often contains one or more SLOs, while additional internal objectives may be stricter than customer-facing commitments. An operational target is not automatically a contractual promise.

In site reliability engineering, objectives may also define an error budget: the amount of unsuccessful service permitted within the target. Operational decisions based on that budget are distinct from contractual remedies.

Typical contents

An SLA commonly addresses:

Component Typical provisions
Service definition Covered products, functions, locations, subscription tiers, and operating hours
Performance commitments Availability, latency, throughput, successful transactions, or support response times
Measurement rules Data sources, observation points, sampling intervals, aggregation, and reporting periods
Responsibilities Provider duties and customer configuration, access, and reporting obligations
Incident handling Severity classifications, communication channels, escalation, and restoration procedures
Exclusions Specified maintenance, customer-caused failures, or other excluded circumstances
Remedies Service credits, corrective-action requirements, or termination rights
Governance Review meetings, dispute handling, amendment procedures, and notice periods

Not every SLA contains all of these elements. Published cloud-service SLAs often concentrate on availability and service credits, while negotiated outsourcing agreements may cover a much wider range of operational responsibilities.

Metrics and measurement

Availability

Availability concerns whether a service is usable under specified conditions. A time-based availability measure can be expressed as:

A=T−DT×100%,A = \frac{T-D}{T}\times 100\%,

where TT is the eligible measurement time and DD is qualifying downtime within that period.

If all time in a 30-day month is included, a 99.9% target permits 43.2 minutes of downtime. That calculation changes if the agreement excludes certain periods or uses a different measurement method.

Some SLAs measure availability through requests rather than elapsed time:

Arequests=NsuccessfulNeligible×100%.A_{\mathrm{requests}} = \frac{N_{\mathrm{successful}}}{N_{\mathrm{eligible}}} \times 100\%.

These measures are not interchangeable. A brief outage during peak traffic can affect many requests while accounting for little elapsed time.

The definition of downtime is therefore central. An agreement may require a minimum continuous failure interval, simultaneous failure of specified components, or unsuccessful responses from a particular endpoint before an outage qualifies.

Latency, throughput, and correctness

Latency measures the time required for an operation. An SLA may specify an average, a maximum, or a percentile threshold. A 99th-percentile latency target characterizes the slower end of the measured request distribution; it does not mean every request completes within that time.

Throughput measures the amount of work completed per unit of time. Its meaning depends on workload definitions, concurrency, payload size, and operating conditions. A throughput commitment without those conditions may be difficult to interpret.

Other indicators include transaction success rates, data freshness, delivery delays, and processing accuracy. For a database service, successful connection establishment is a different commitment from successful completion of a transaction.

Support and recovery

Support commitments distinguish among:

  • Response time: time until a qualifying acknowledgment or engagement.
  • Restoration time: time until usable service is restored, potentially through a workaround.
  • Resolution time: time until the underlying issue is corrected.

A promise to respond within one hour is not a promise to restore service within one hour. Definitions of incident severity and support hours determine when such commitments apply.

Recovery provisions may include a recovery time objective (RTO), describing the targeted time to restore operations after disruption, and a recovery point objective (RPO), describing the acceptable age of data to which operations can be restored. These concern recovery capability rather than routine availability alone. The existence of data backups does not, by itself, establish either objective.

Measurement boundaries and exclusions

An SLA must identify where performance is observed. A provider may measure availability at its service endpoint, whereas a customer experiences the complete path through its own network, the Internet, authentication systems, and application dependencies. Both measurements can be accurate while describing different service boundaries.

Important measurement choices include:

  • Whether monitoring is provider-operated, customer-operated, or independent.
  • Whether results are calculated per account, region, instance, endpoint, or service.
  • How missing observations, retries, and partial failures are treated.
  • Whether aggregation is weighted by time, request volume, or another quantity.
  • Which requests and operating configurations are eligible.

Possible exclusions include scheduled maintenance, unsupported configurations, customer actions, and events outside the provider’s defined responsibility. Their effect depends on their wording. A maintenance window may be excluded, included, or subject to a separate limit.

Remedies and administration

A common remedy is a service credit, usually calculated as a percentage of eligible charges and applied to a future bill. Credit schedules may increase as measured performance falls below successive thresholds.

Credits are not necessarily cash refunds and need not correspond to the customer’s actual losses. Agreements may impose claim deadlines, evidence requirements, credit caps, and minimum eligibility conditions. Some identify credits as the exclusive remedy for specified failures; others provide additional rights for repeated or prolonged breaches.

SLA administration also includes performance reports, incident reviews, escalation procedures, and periodic reassessment. Effective administration requires consistent records of measurements, exclusions, and incident timelines, since a numerical target alone does not resolve disputes about whether a breach occurred.

Applications and limitations

SLAs are used for network connectivity, hosted applications, managed cybersecurity services, technical support, and software as a service. Their operational importance has grown as organizations depend on services operated outside their direct control.

An SLA for one component does not automatically establish an end-to-end commitment. Applications built from multiple services can fail because of dependencies, integration errors, or correlated outages. Under the simplifying assumptions that all components must work and their failures are independent, combined availability is:

Acombined=∏i=1nAi,A_{\mathrm{combined}}=\prod_{i=1}^{n} A_i,

where each AiA_i is expressed as a fraction. Real systems may violate these assumptions through shared infrastructure, redundancy, or common failure causes.

Aggregate metrics can also conceal uneven experiences: a service may meet a global target while one region or customer group suffers substantial disruption. Thresholds and exclusions can create incentives to optimize the reported indicator rather than the underlying user experience.

Finally, an availability commitment is not a guarantee of confidentiality, data integrity, or regulatory compliance. Those obligations require separate definitions. An SLA specifies selected aspects of service accountability, not every dimension of service quality.