This article contains no advertising or affiliate links. We rechecked official Google SRE, AWS public SLA, and Uptime Institute materials and recalculated downtime on July 29, 2026. For a particular service, use the latest official contract terms.
Availability cannot be compared from a number such as 99.9% alone.The same number means different things depending on what counts as available, the period and locations measured, and whether planned maintenance or user-caused failures are excluded.
99.9% corresponds to 43 minutes 12 seconds in 30 days; read it alongside the measurement definition
Availability is the proportion of time or requests for which a service is usable. Distinguish an observed SLI, an SLO target, and a contractual SLA; align the period, scope, success criteria, and exclusions when evaluating them.
| Meaning of availability | The proportion of a defined period or valid requests in which users could use the service successfully |
|---|---|
| 99.9% over 30 days | 43 min 12 s of equivalent unavailable time |
| 99.9% over 365 days | 8 h 45 min 36 s of equivalent unavailable time |
| Conditions for comparison | Align measurement period, scope, success criteria, exclusions, partial-failure aggregation, and observation points |
| SLI, SLO, and SLA | An SLI is a measured indicator, an SLO is a target, and an SLA is a contractual agreement including consequences for missing it |
| Last checked | July 29, 2026. Use each service's latest official contract terms |
Server availability is the proportion in which users could use a service
In server operations, availability or uptime is the proportion of a defined period or valid requests for which the service was available.
The Japanese term for operational rate can also mean utilization of manufacturing equipment or CPU resources. This article limits the term to availability of hosted servers, cloud services, websites, and web applications.
What matters is whether users can complete necessary functions, not merely whether the server is powered on. Even if ping succeeds, failures in DNS, TLS certificates, databases, applications, or external APIs may prevent login or purchases, making the service unavailable from a user perspective.
Time-based availability
The proportion of a period such as a month during which the service was usable. It is easy to understand but weights quiet nights and busy daytime equally.
Request-based availability
The proportion of valid requests that succeeded. It can better reflect partial failures and changing traffic volumes.
User-perspective availability
Check whether important journeys such as login, search, saving, and checkout can be completed, not just whether a page loads.
System-wide availability
Evaluate the result across necessary components including servers, DNS, CDN, databases, authentication, and external APIs.
Uptime and availability are not always exact synonyms
Hosting descriptions use them similarly, but uptime can mean how long equipment or a process runs. For service quality, check the definition of availability, emphasizing whether users can complete operations.
Two ways to calculate server availability
Google SRE describes time-based and successful-request-based approaches. In either case, first define what the numerator and denominator include.
Thirty days is 43,200 minutes. With 43.2 minutes unavailable, 43,156.8 ÷ 43,200 × 100 = 99.9%.
If 999,000 of 1,000,000 valid requests succeed, 999,000 ÷ 1,000,000 × 100 = 99.9%.
Calculating 99.9% over 30 days
Convert the measurement period to minutes
30 days × 24 hours × 60 minutes = 43,200 minutes.
Aggregate unavailable time according to the definition
Add downtime from monitoring records. Here it is 43.2 minutes.
Find available time
43,200 − 43.2 = 43,156.8 minutes.
Convert to a percentage
43,156.8 ÷ 43,200 × 100 = 99.9%.
Reverse calculation: allowable unavailable time
Measurement period × (1 − target availability) gives the result. At 99.9% over 30 days, 43,200 minutes × 0.001 = 43.2 minutes. This is the equivalent limit derived from a target, not planned or actual downtime.
Downtime from 99% to 99.999%
We use 30-day months and 365-day years, calculating period × (1 − availability). Calendar-month SLAs have different allowances for 28, 29, and 31 days.
The table scrolls horizontally.
| Availability | 30 days (43,200 minutes) | 365 days (525,600 minutes) | Unavailable proportion |
|---|---|---|---|
| 99% | 7 h 12 min | 3 d 15 h 36 min | 1% |
| 99.5% | 3 h 36 min | 1 d 19 h 48 min | 0.5% |
| 99.9% | 43 min 12 s | 8 h 45 min 36 s | 0.1% |
| 99.95% | 21 min 36 s | 4 h 22 min 48 s | 0.05% |
| 99.99% | 4 min 19.2 s | 52 min 33.6 s | 0.01% |
| 99.999% | 25.92 s | 5 min 15.36 s | 0.001% |
Quick downtime calculator
Enter availability and a period to calculate the equivalent maximum unavailable time.
Equivalent error budget: 0.1%
Each additional nine reduces allowable downtime to about one tenth
99.9% and 99.99% differ by 0.09 percentage points, but equivalent unavailable time over 30 days falls from 43 minutes 12 seconds to 4 minutes 19.2 seconds. Higher targets require greater investment in monitoring accuracy, redundancy, change management, and automated recovery.
Six conditions to fix before measuring availability
Availability is defined by its measurement specification. Whether comparing providers or operating internally, do not directly compare numbers with different conditions.
Measurement period
Specify a calendar month, rolling 30 days, quarter, 365 days, or another period. Simply averaging monthly SLA values into a yearly value hides differences in month length and request volume.
Scope
Define the target: one server, region, account, whole website, or particular API. Check whether DNS and the management console are excluded.
Success criteria
Define HTTP status, response content, completion, and response time. Decide whether a 200 response with broken content or extreme latency counts as success.
Exclusions
Check planned maintenance, customer configuration, upstream providers, force majeure, beta features, and similar exclusions. Broader exclusions increase the gap from experienced downtime.
Partial-failure aggregation
Decide whether time when only 10% of users fail counts as total downtime or is weighted by failed-request rate. Important journeys can also have separate SLIs.
Observation points and evidence
Decide whether provider monitoring, external monitoring from multiple regions, or user logs are authoritative, and retain times, locations, errors, and durations.
In 2026, event-based measures for important journeys are practical
When only some regions or APIs fail, successful valid events ÷ total valid events can represent impact more continuously than downtime alone. However, large numbers of page views can hide a low-volume payment failure, so separate SLIs by user journey.
One server's SLA is not the availability of the entire site
If using a site requires DNS, CDN, a load balancer, application, database, authentication, and payment, one component's SLA alone does not determine overall availability.
Simplified case: all are required and failures are independent: 0.999 × 0.999 × 0.999
Approximately 99.7003%In this simplified example, unavailable time over 30 days is about 2 hours 9 minutes. In reality, correlated failures from shared networks or configuration changes and mitigation through caches or alternative paths mean multiplication alone cannot predict the outcome. Still, it shows that components each at 99.9% do not necessarily make a site 99.9% available.
Do not count an unusably slow service as available
Responses just before a timeout or a page that cannot be used after displaying are practical failures even with HTTP 200. Define success and sufficient speed for each important journey, and also use latency SLOs such as 95% of requests completing within a defined time.
Monitoring only the server can miss user-facing failures
At minimum, periodically exercise important journeys such as the home page, login, search, saving, and checkout externally. Internal metrics help diagnose causes; synthetic monitoring helps identify user impact. Combine both.
Availability, SLI, SLO, SLA, and related metrics
Related terms have distinct roles: measurement, target, contract, and design property. Availability can be an SLI but is not itself a target or contract.
The table scrolls horizontally.
| Term | Role | Relationship to availability | Example |
|---|---|---|---|
| SLI Measure | A quantitative indicator of service level | Availability, success rate, latency, and correctness can be SLIs. | Purchase success rate of 99.93% |
| SLO Target | A target value or range for an SLI | The level to achieve, not the measured value. | At least 99.9% over 30 days |
| SLA Agreement | An agreement defining the level and consequences of not meeting it | Read the formula, exclusions, remedies, and claim requirements. | A partial service credit if the target is missed |
| Error budget | The failures permitted by an SLO | The basic form is 100% − SLO: 0.1% for 99.9%. | Equivalent to 43 min 12 s over 30 days |
| Reliability | The property of performing expected functions without failure | The same total downtime can yield the same availability despite different failure counts and recovery times. | Failure frequency and recoverability |
| Redundancy Design | Duplicating components or paths | A way to improve availability, though failover and common-cause failures remain possible. | Multiple availability zones and duplication |
| Utilization | The proportion of CPU, network, or equipment in use | Higher is not always better and may indicate pressure. | CPU utilization of 75% |
| Durability | The property of not losing stored data | Data may be retained even if temporarily inaccessible. | Object durability |
| MTBF / MTTR | Time between failures and time to recover or repair | A simple steady-state model estimates availability as MTBF ÷ (MTBF + MTTR), but this does not replace a service SLI. | Reduce recovery time |
Do not convert data-center tiers into fixed availability figures
Uptime Institute says it removed references to expected annual downtime from the Tier Standard in 2009, and the current Tier Standard: Topology does not assign availability predictions to tier levels. Tier addresses facility infrastructure maintainability and fault tolerance, distinct from application SLAs and measured results.
Five steps to compare 99.9% SLAs
Do not choose an SLA by its number alone. Even AWS's public SLAs differ by service in measurement unit, unavailability definition, credits, exclusions, and claim conditions. Check the latest terms for the service you contract.
- Check the covered service and unitCheck whether coverage applies to a region, instance, account, or endpoint. It may not cover the entire website.
- Check the formula and measurement periodRead whether it uses calendar months or rolling periods, time or error rate, and provider logs or external monitoring.
- Check the definition of unavailability and exclusionsCheck whether planned maintenance, customer configuration, external networks, force majeure, or beta features are excluded.
- Check remedies and claim requirementsCheck whether credits are automatic or require a claim, deadlines, evidence, eligible charges, and credit caps.
- Separately check measured history and your own recovery measuresAn SLA is not historical performance. Check official status, external monitoring, incident reports, backups, and failover tests.
A service credit may not fully compensate business losses
Many public SLAs provide credits against charges for the covered period. Lost sales, idle employees, and reputational harm are not necessarily compensated automatically. Separately plan for unacceptable impact using redundancy, backups, alternatives, recovery objectives, and individual contracts.
Short SLA checklist
- Is the denominator for 99.9% time, requests, or five-minute intervals?
- Does unavailability include only failed connections, or also error responses and latency?
- Must the customer configure multiple availability zones or regions?
- Are planned maintenance, third-party services, or customer configuration excluded?
- What are the claim deadline and required logs, times, and resource IDs?
Eight practices for higher server availability
High availability takes more than duplicating servers. Combine design, monitoring, change, recovery, and organizational operations.
Reduce single points of failure
Consider load balancers, multiple zones, redundant databases, and redundant DNS, and test that failover actually works.
Monitor important journeys externally
Monitor pages, login, search, saving, and checkout from multiple locations, checking content and response time.
Operate SLOs and error budgets
Set targets and use rapid budget consumption to guide decisions such as pausing changes or prioritizing reliability.
Make changes safely
Use small releases, staged rollout, rollback, and configuration review to limit the scope and recovery time of change-related failures.
Keep capacity headroom
Monitor CPU, memory, connections, queues, and database load, and plan autoscaling and limits for spikes.
Test backups and restoration
Do more than verify that backups exist: regularly test recovery time, restored consistency, permissions, and procedures.
Isolate dependency failures
Use timeouts, retry limits, circuit breakers, queues, and degraded operation to prevent cascading failures.
Practice incident response
Assign owners, communication paths, user notices, and failover procedures, then track prevention and detection improvements in post-incident reviews.
Targeting 100% sharply increases cost and slows change
Google SRE explains that always requiring 100% is unrealistic and can lead to overly conservative design and slower development. Set an appropriate SLO and error budget based on business impact and investment.
Frequently asked questions about server availability
Does 99.9% mean about 52 minutes of downtime per year?
No. At 99.9% over 365 days, the equivalent is 8 hours 45 minutes 36 seconds. About 52 minutes 33.6 seconds corresponds to 99.99%. Both are arithmetic conversions and do not apply SLA exclusions.
How different are 99.9% and 99.99%?
Over 30 days, 99.9% corresponds to 43 minutes 12 seconds and 99.99% to 4 minutes 19.2 seconds, about one tenth as much. SLAs with different scopes or exclusions cannot be compared by number alone.
Is planned maintenance excluded from availability?
Not universally. Follow the formula in your internal SLI or contractual SLA. Recording user-experienced unavailability separately from contractual unavailability allows both experience and remedy eligibility to be tracked.
Does a 99.9% SLA guarantee actual availability?
An SLA defines service levels and consequences of missing them; it is not a historical measurement. Some contracts apply service credits if the level falls below the threshold under their scope and exclusions.
Are uptime and availability the same?
Hosting providers often use them similarly. Technically, uptime can mean equipment or process running time, while availability emphasizes whether users can use required functions. Check definitions when comparing.
If ping succeeds, is the service available?
Ping alone is insufficient for an important service. It can succeed while the web returns 500 errors, login fails, or saves fail. Monitor important user journeys externally and define success and response-time conditions.
Does Tier III determine a data center's availability percentage?
It cannot be converted into a fixed percentage. Uptime Institute's current Tier Standard: Topology does not assign availability predictions to tier levels. Evaluate facility design, operations, networks, applications, SLAs, and actual results separately.
Primary and official sources reviewed
We checked definitions, calculations, SLOs, error budgets, public SLAs, and tiers against these sources. Service-specific terms change, so check the latest version before contracting and when making a claim.
- Google SRE Book: Availability Table — availability, equivalent downtime, and useful aggregate indicators for partial failure
- Google SRE Book: Embracing Risk — time-based and aggregate availability and error budgets
- Google SRE Book: Service Level Objectives — SLI, SLO, SLA, and cautions about 100% targets
- Google SRE Workbook: Implementing SLOs — event-based SLIs and error-budget operations
- Amazon S3 Service Level Agreement — examples of monthly availability, service credits, and terms
- Amazon Compute Service Level Agreement — examples of region- and instance-level measurement units
- Uptime Institute: Explaining the Tier Classification System — why tiers are not tied to fixed availability
Summary: check what and how 99.9% measures
Server availability is the proportion for which a service was usable. At 99.9%, unavailable time is equivalent to 43 minutes 12 seconds over 30 days or 8 hours 45 minutes 36 seconds over 365 days. Comparisons require aligned periods, scope, success criteria, exclusions, partial failures, and observation points.
Distinguish measured SLIs, target SLOs, and contractual SLAs, and check site-wide dependencies, latency, actual history, and recovery procedures. Continuously measure whether users can complete important actions, not just a number.