Availability measures whether a workload performs its agreed function over a stated period. Define the business target, distinguish it from service SLA commitments, and account for required dependencies and independent alternatives.
Service SLAs and workload targets
A workload availabilityThe percentage of a measurement period during which a workload performs its agreed function successfully. target states how often the application must perform its agreed function. A service-level agreementA service commitment that defines measurement and remedies if it is not met., or SLA, defines a service commitment and its remedy. Set the workload target from business needs and evaluate the services it depends on. AWS availability guidance · SLA and dependency guidance
How it works
- Define the workload function and measurement period. Availability measures whether the function succeeds during that period.
- Define recovery limits separately. RTOThe maximum acceptable delay from service interruption to restoration., the recovery time objective, limits the delay until service restoration. RPOThe maximum acceptable time since the last recoverable data point., the recovery point objective, limits the age of the recoverable data and therefore the acceptable data loss.
- Read each dependency SLA for deployment conditions and its definition of unavailability. An Availability ZoneAn isolated location within an AWS Region., or AZ, is an isolated location within an AWS Region.
- Choose a target the business needs and can support. Higher targets typically require more cost, testing and automated recovery.
DNSThe system that returns information about resources associated with domain names., the Domain Name System, answers queries about domain names. A name serverA server that answers DNS queries with information about domain names. answers these queries; an authoritative name server uses the records it holds for the domain.
| Measure or commitment | Scope | What to check |
|---|---|---|
| Workload availability target | Successful operation of the whole workload | Function and measurement period |
| EC2 Region-Level SLA | Qualifying EC2 deployment | 99.99% monthly; instances across at least two AZs, with the single-AZ Region exception |
| EC2 Instance-Level SLA | Individual EC2 instance | 99.5% monthly |
| Route 53 Authoritative DNSThe DNS service that answers using the records held for a domain. SLA | Hosted-zone DNS responses | Assigned name servers and SLA exclusions |
EC2 Region-Level unavailability requires all running instances in the qualifying AZs to have no external connectivity concurrently. A qualifying service-credit claim goes through AWS Support and is due by the end of the second billing cycle after the incident.
A Route 53 hosted zoneA container for DNS records describing traffic routing for a domain and its subdomains. is a container of DNS records that describe where a domain’s traffic should go. The Authoritative DNS SLA treats a minute as unavailable only when all four assigned virtual name servers fail to answer all DNS queries throughout that minute. Other Route 53 features, including its API and console, are excluded. Hosted zones · Route 53 SLA
Exam cues
- If the question says one EC2 instance must qualify for the Region-Level commitment, think deployment across AZs, because the Region-Level SLA has an explicit deployment condition.
- If the question says an application failed while DNS continued answering, think workload availability, because the Route 53 SLA measures hosted-zone DNS responses.
- If the question says restore service quickly and lose little data, think separate RTO and RPO values, because restoration delay and recoverable data age measure different effects.
Remember
- EC2 Region-Level SLA: 99.99% monthly, qualifying concurrent deployment across at least two AZs. If a Region has only one AZ, the SLA specifies at least two Regions.
- EC2 Instance-Level SLA: 99.5% monthly.
- Higher availability targets increase engineering and operational demands.
Trap: A dependency’s SLA percentage is not a prediction of the workload’s actual downtime. Measure the workload function and account for its dependencies.
From the AWS docs (16)
- Workload availability measures the percentage of a stated period during which the workload performs its agreed function successfully.
- An SLA defines a service commitment, its measurement and remedies if the commitment is not met.
- Higher availability targets typically increase cost and require more extensive failure testing and recovery automation.
- Recovery time objective is the maximum acceptable delay between service interruption and restoration.
- Recovery point objective is the maximum acceptable time since the last data recovery point, expressing acceptable data loss.
- The EC2 Region-Level SLA commits to at least 99.99% monthly uptime for running instances concurrently deployed across two or more AZs in a Region, or at least two Regions if a Region has only one AZ.
- The EC2 Instance-Level SLA commits to at least 99.5% uptime for an individual instance in a monthly billing cycle.
- EC2 Region-Level unavailability means all running instances in the qualifying AZs concurrently have no external connectivity.
- EC2 SLA service credits require a qualifying claim through AWS Support; credit requests are due by the end of the second billing cycle after the incident.
- An Availability Zone is an isolated location within an AWS Region.
- The Route 53 SLA covers Authoritative DNS per hosted zone; a hosted zone is unavailable for a minute only when all four assigned virtual name servers fail to answer all DNS queries throughout that minute.
- The Route 53 Authoritative DNS SLA excludes other Route 53 features, including API and console availability.
- A hosted zone is a container for DNS records that describe traffic routing for a domain and its subdomains.
- SLA availability figures are not a guarantee of actual downtime; workloads and dependencies can exceed their stated SLA availability.
- DNS means Domain Name System, which returns information about resources associated with domain names.
- An authoritative name server answers DNS queries using definitive information for its part of DNS; Route 53 name servers use hosted-zone records for their domains.
Downtime budgets and nines
A downtime budgetThe unavailability compatible with a target over its measurement period. is the amount of unavailability compatible with an availabilityThe percentage of a measurement period during which a workload performs its agreed function successfully. target over a stated period. NinesShorthand for an availability percentage, such as five nines meaning 99.999%. are shorthand for availability percentages: five nines means 99.999%.
How it works
- Choose the measurement period and define when the workload is usable. Time-based availability is usable time divided by total time.
- Express the target as a fraction. Rearranging that ratio gives
allowed downtime = period × (1 - target availability). - Compare interruptions against that budget using the same period. Under the strict interpretation, scheduled and unscheduled interruptions both reduce availability.
- For request-based measurement, divide successful responses by valid requests. AWS describes one-minute or five-minute intervals, averaged to obtain monthly uptime; an interval with no requests counts as fully available.
Exam cues
- If the question says roughly 50 minutes of downtime per year, think 99.99%, because AWS lists about 52 minutes at that tier.
- If the question says roughly five minutes per year, think five nines, because that is the annual budget listed for 99.999%.
- If the question says remove maintenance from the calculation, think the agreed workload measurement and AWS’s advice to include it, because users can need the service during planned work.
Remember
| Availability target | Approximate annual downtime from AWS’s table |
|---|---|
| 99% | 3 days 15 hours |
| 99.9% | 8 hours 45 minutes |
| 99.95% | 4 hours 22 minutes |
| 99.99% | 52 minutes |
| 99.999% | 5 minutes |
These are rounded annual values. Use the actual measurement period when calculating a budget.
Trap: An annual downtime table does not give a monthly allowance. Calculate from the period named in the requirement.
From the AWS docs (11)
- Time-based availability is usable time divided by total time over a specified measurement period.
- Rearranging the time-based availability formula gives allowed downtime as total time multiplied by one minus target availability, using availability as a fraction.
- Five nines means 99.999% availability.
- AWS lists approximately 3 days 15 hours of annual unavailability for a 99% target.
- AWS lists approximately 8 hours 45 minutes of annual unavailability for a 99.9% target.
- AWS lists approximately 4 hours 22 minutes of annual unavailability for a 99.95% target.
- AWS lists approximately 52 minutes of annual unavailability for a 99.99% target.
- AWS lists approximately 5 minutes of annual unavailability for a 99.999% target.
- The strict availability interpretation includes scheduled and unscheduled interruptions; AWS advises against excluding scheduled downtime.
- Availability can be measured as successful responses divided by valid requests.
- AWS describes measuring request-based availability in one-minute or five-minute periods and averaging periods for monthly uptime; an interval without requests counts as 100% available.
Dependency and redundancy calculations
AvailabilityThe percentage of a measurement period during which a workload performs its agreed function successfully. calculations depend on what must work together. A hard dependencyA component whose failure stops the workload from performing its function. must work for the workload to perform its function. A soft dependencyA component whose failure the application can tolerate or compensate for. can fail while the application compensates. Redundant components are alternatives that can keep the function available when another component fails. AWS dependency model · RedundancyAlternative components that keep a workload operating when another component fails. model
How it works
- Identify required components, including the application itself. For independent hard dependencies, multiply their availability fractions.
- Identify alternatives that can each handle the workload alone. Calculate the probability that all alternatives fail by multiplying their independent failure probabilities, then subtract that value from one.
- Check the independence assumption. An Availability ZoneAn isolated location within an AWS Region. is a shared failure boundaryA scope within which multiple resources can fail together.: its resources can fail together. Redundancy needs failure detection, work shifting and enough healthy capacity.
- If a component has no published availability guidance, estimate it from mean time between failuresThe average interval between component failures., or MTBF, and mean time to recoverThe average time needed to recover after failure., or MTTR. Use
MTBF / (MTBF + MTTR), with both durations in the same unit.
| Arrangement | Calculation with availability fractions | Example |
|---|---|---|
| Both independent components required | A × B |
0.999 × 0.999 = 99.8001% |
| Either independent component sufficient | 1 - ((1 - A) × (1 - B)) |
Two 99.9% alternatives give 99.9999% |
| Estimate from failure and recovery duration | MTBF / (MTBF + MTTR) |
150 days and 1 hour give about 99.97% |
Exam cues
- If the question says every independent dependency must work, think multiplication, because failure of any required component stops the workload function.
- If the question says either independent replica can serve all work, think one minus the product of failure probabilities, because all replicas must fail to stop service.
- If the question says both replicas share one AZ, think a shared failure boundary, because an AZ failure can affect both together.
Remember
- Three independent required components at 99.99% each give approximately 99.97% theoretical availability.
- Two independent alternatives at 99.9% each give 99.9999% theoretical availability.
- The ninesShorthand for an availability percentage, such as five nines meaning 99.999%. shortcut adds nines only for independent redundant alternatives whose percentages consist entirely of nines.
- Calculations based on SLAA service commitment that defines measurement and remedies if it is not met. or SLOA target level of service used as an availability goal. figures are rough estimates. An SLO, or service-level objective, is an availability goal.
Trap: Two copies do not automatically justify the independent-redundancy formula. Shared failure boundaries, failure detection and surviving capacity must fit the model.
Box details (5)
- Component A
A required component belongs in the product of availabilities.
- Component B
Failure of either required component interrupts this workload function.
- Client
The redundant model assumes either alternative can perform the workload function.
- Replica A
Both independent alternatives must fail to interrupt the redundant function.
- Replica B
Failure detection, work shifting and healthy capacity are needed to use this alternative.
From the AWS docs (14)
- A hard dependency is a component whose failure prevents the workload from performing its function.
- A soft dependency can fail while the application compensates for or tolerates that failure.
- For independent required components, theoretical workload availability is the product of their availabilities.
- Using the required-component formula, two independent components each available 99.9% give 0.999 multiplied by 0.999, or 99.8001% availability.
- AWS gives three independent required components at 99.99% each as approximately 99.97% theoretical workload availability.
- If independent redundant components can each serve the workload alone, theoretical availability is one minus the product of their failure probabilities.
- Two independent redundant components at 99.9% each yield 99.9999% theoretical availability.
- For independent redundant components whose availability percentages contain only nines, summing the count of nines gives the resulting nines count.
- MTBF means mean time between failures and MTTR means mean time to recover.
- An availability estimate can be calculated as MTBF divided by the sum of MTBF and MTTR.
- AWS gives MTBF of 150 days and MTTR of 1 hour as an availability estimate of about 99.97%.
- Availability calculations using SLA or SLO figures are rough estimates, not a guarantee of actual downtime.
- Resources in one Availability Zone can fail together, so AZ-level sparing must account for the loss of all resources in that AZ.
- Using redundant spares requires detecting failures, shifting work to healthy resources and managing workload capacity.
Failure domains and spare capacity
A failure domainA scope within which multiple resources can fail together. is a location whose resources can fail together, such as an Availability ZoneAn isolated location within an AWS Region. (AZ). SpareAlternative components that keep a workload operating when another component fails. capacity is already available capacity that can take over lost work. Static stabilityThe ability to keep operating through failure using resources and configuration prepared beforehand. means preparing the surviving locations to serve the workload before a failure occurs.
How it works
- Choose the failure to tolerate: an instance, an AZ, or a Region. Distribute resources across the relevant boundary. Multiple instances in one AZ still share that location’s risk.
- Pre-provision enough capacity for the surviving locations. This avoids making immediate recovery depend on the control planeThe part of a service that creates, changes, or deletes resources., which creates or changes resources. The data planeThe part of a service that carries out work using existing resources. is the existing resources’ working functionality.
- Shift traffic to healthy instances using health checks. Auto Scaling can repair lost capacity afterward.
- Check storage, quota headroomUnused quota that allows replacement resources to coexist with resources affected by failure., and subnetA range of network addresses allocated within a virtual private cloud. address space. A quota is an account usage limit; failed resources may still count while replacements launch. A subnet is a network address range whose free addresses must accommodate growth.
- Test the expected load with faults injected. Recovery must work across the application and its dependencies. Quota planning · Network planning · Resilience testing
| Storage choice | Interface | Failure-domain consideration |
|---|---|---|
| EBS | Block storageStorage presented to an instance as a device used like a disk., used like an attached disk | Volume and instance must share an AZ |
| EFS Regional / One Zone | Shared files | Regional spans AZs; One Zone can lose data with its AZ |
| S3 Standard / One Zone-IA | Objects | Standard spans at least three AZs; One Zone-IA uses one |
Block storage · File system types · Object storage
Exam cues
- If the question says survive an AZ failure without waiting for new instances, think static stability, because surviving capacity is provisioned in advance.
- If the question says replacement launches hit quotas during failover, think overlap headroom, because inaccessible instances may still consume quota.
- If the question says compute spans AZs but storage remains in one, think a storage failure domain, because an EBS volume cannot attach across AZs.
- If the question says survive a Region failure, think an explicit multi-Region design, because EC2 does not automatically replicate resources between Regions.
Remember
- Check surviving capacity, storage placement, quotas, and free network addresses together.
- S3 Standard spans at least three AZs; S3 One Zone-IA uses one.
- EFS Regional and One Zone have different failure scopes.
- Test recovery under expected load.
Trap: A second AZ is not proof of sufficient failover capacity. Waiting to launch resources adds a control-plane dependency during recovery.
From the AWS docs (12)
- An Availability Zone is an isolated location within a Region; instances in the same location can become unavailable together.
- Regions are isolated from one another, and EC2 resources are not automatically replicated between Regions.
- Static stability for compute means provisioning enough capacity before failure for the surviving locations to handle the workload.
- Pre-provisioning removes the need to launch replacement instances through the EC2 control plane as part of immediate recovery.
- The control plane creates or changes resources; the data plane provides the existing resources' working functionality.
- Load balancer or Route 53 health checks can shift traffic from impaired instances; Auto Scaling can replace lost capacity afterward.
- Inaccessible resources can still consume quotas, so failover quotas must allow failed resources and their replacements to overlap.
- Subnet address allocation must accommodate expansion and high availability.
- Fault injection and expected-load testing exercise recovery behavior that architecture alone does not prove.
- EBS provides block storage, and an attached volume and EC2 instance must be in the same Availability Zone.
- EFS Regional stores file data across multiple AZs; EFS One Zone stores it in one AZ and can lose data if that AZ is lost.
- S3 Standard stores objects across at least three AZs; S3 One Zone-IA stores objects within one AZ.
Flashcards
Nothing to review here right now. Pick other topics or another filter.
·
All 21 cards as a list
How does an AWS service SLA differ from a workload availability target?
An SLA defines a service commitment, measurement and remedies. A workload target concerns successful operation of the whole workload over its measurement period. EC2 SLA credits do not establish application availability.
What is a tradeoff when a business sets a higher workload availability target on AWS?
Higher targets typically increase cost and require more failure testing and recovery automation.
What does an AWS workload availability measure describe?
The percentage of a stated period when the workload performs its agreed function successfully.
Which workload objective limits service restoration time, and which limits the age of recoverable data?
RTO limits the delay until restoration. RPO limits the time since the last data recovery point and therefore acceptable data loss.
What does an AWS service level agreement (SLA) describe?
It defines a service commitment, how it is measured, and remedies if the commitment is not met.
A business permits roughly 50 minutes of yearly unavailability for its order system. Which availability tier is closest?
99.99%. AWS lists about 52 minutes of annual unavailability at that target.
How do you turn a workload availability target into a downtime budget?
Multiply the measurement period by one minus the target expressed as a fraction. Keep the target and budget on the same time basis.
A workload must meet five nines annually. How much annual unavailability does AWS list for that tier?
About 5 minutes for 99.999% availability.
A team excludes planned maintenance from its workload availability report to meet its target. Does AWS recommend this approach?
No. AWS advises against excluding scheduled downtime, because users may still need the workload during maintenance.
A workload requires two independent components, each 99.9% available. What is its theoretical availability?
99.8001%. Multiply 0.999 by 0.999 because both components must work.
A workload can use either of two independent redundant components, each 99.9% available. What is its theoretical availability?
99.9999%. Calculate 1 - (0.001 × 0.001), because both independent components must fail to stop the workload.
How can a workload designer estimate dependency availability from MTBF and MTTR?
Use MTBF / (MTBF + MTTR), with both durations in the same unit. AWS gives 150 days MTBF and 1 hour MTTR as about 99.97%.
An application tolerates failure of an optional dependency and still performs its agreed function. Should that dependency be treated as required in the serial availability product?
No. It is a soft dependency for that function because the application compensates for its failure.
Two EC2 replicas share one Availability Zone. Can a designer assume their AZ failure risks are independent?
No. Resources in one AZ can fail together. Model the AZ as a shared failure boundary and add redundancy across that boundary.
Does multiplying the SLA percentages of a workload's dependencies predict its actual annual downtime?
No. It gives a rough theoretical estimate. SLA figures do not guarantee actual downtime and dependencies can exceed them.
What does static stability mean for a multi-AZ EC2 workload?
Pre-provision enough capacity in surviving AZs to serve the workload without waiting for new instance launches.
An EC2 application spans several AZs but needs new instances before it can serve normal load after losing an AZ. What resilience improvement is needed?
Provision spare capacity before the failure. A multi-AZ layout alone does not remove the recovery dependency on instance launches.
An EC2 failover cannot create replacement instances because inaccessible instances still consume the account quota. What should the design allow?
Quota headroom for both inaccessible resources and replacements at the same time.
An EC2 application has healthy replicas in another AZ but its data is on an EBS volume in the failed AZ. Can the replicas attach that volume across AZs?
No. EBS volumes attach to instances in their own AZ. The storage design must also meet the failure-domain requirement.
Which EFS file system type stores shared file data across AZs, and how does S3 Standard compare?
EFS Regional stores files across multiple AZs. S3 Standard stores objects across at least three AZs. EFS One Zone and S3 One Zone-IA keep data in one AZ.
An EC2 service already has enough surviving capacity during an AZ outage. How should traffic shifting and capacity repair relate?
Shift traffic to healthy instances using health checks, then let Auto Scaling restore lost capacity. Immediate service recovery need not wait for new instances.