3  Cloud Service Models for IoT

reference-architectures
cloud
computing
service

3.1 Start With the Workload Duty

A cloud service model is a duty split. A team may own the application logic, the runtime, the operating system, the broker, or only the data contract, and each split changes what the team can fix during an IoT incident.

Start simple: pick one workload, such as alarm routing or firmware distribution, and ask who patches it, scales it, observes it, and recovers it. The right model is the one whose duty split matches that evidence. ## Service Models Move Duties {#overview-cloudsvcdepth-boundaries .depth-l0}

Cloud service models are responsibility boundaries. They describe which layers the IoT team operates directly and which layers a service provider operates under a contract. A model is not automatically mature, secure, cheap, or scalable. It becomes useful only when the team can name the device, data, runtime, access, recovery, and support responsibilities that remain on each side of the boundary.

IaaS gives the team more control over infrastructure and runtime choices. PaaS delegates more platform operation while the team still owns device behavior, application logic, data rules, and access policy. SaaS delivers a finished application surface, but the team still owns configuration, onboarding quality, user roles, integration contracts, export tests, and exit planning.

Cloud service model responsibility matrix showing traditional, IaaS, PaaS, and SaaS ownership distribution from networking and servers through virtualization, operating system, runtime, data, and applications.
Choosing IaaS, PaaS, or SaaS changes the operating boundary, not the need for proof.

For a cold-chain monitoring product, IaaS might mean running Mosquitto, PostgreSQL, and dashboard workers on virtual machines that the team patches, backs up, monitors, and restores. PaaS might mean using AWS IoT Core or Azure IoT Hub for device identity, MQTT ingestion, routing rules, and managed queues while the team still owns topic design, payload validation, alert policy, data retention, and application release behavior. SaaS might mean using a finished fleet dashboard or work-order surface, but the team still owns device enrollment quality, role assignment, export rehearsals, integration tests, and the exit path if the workflow no longer fits.

The model is therefore a statement about accountable work. It should say who rotates gateway certificates, who reviews broker topic permissions, who restores historical telemetry, who changes alert thresholds, who answers operator tickets, and who proves that a site can keep alarming during provider or network degradation. Those responsibilities remain visible even when the provider operates more of the stack.

IaaS

The team gets flexible infrastructure and usually owns operating systems, runtime hardening, broker deployment, patching, scaling, backup, monitoring, and recovery drills.

PaaS

The provider manages more runtime capability, while the team owns application behavior, device model, data classification, policy, integration, observability, and release proof.

SaaS

The provider supplies the application surface, while the team owns configuration, user access, data quality, device onboarding, integration fit, exports, and exit evidence.

Design rule: select the model that best fits workload behavior and operating capacity. Do not choose a model because its label sounds modern or because another project used it.

3.2 Match Models to Workloads

Start with the IoT workload, not the cloud label. A telemetry archive, an alarm route, a command channel, a firmware workflow, and an operator dashboard can have different service-model choices inside the same product. The selection is strongest when it documents what the workload must do, which team can operate each layer, what the provider is expected to do, and which change signal will reopen the choice.

IoT service-model fit map linking workload intent, device identity, protocol needs, data ownership, operations capacity, customization, portability, and proof records to IaaS, PaaS, SaaS, or hybrid shapes.
The selection path starts with device and operating facts, then maps those facts to IaaS, PaaS, SaaS, or a hybrid boundary.
Decision Area
What To Ask
Likely Pressure
Proof To Keep
Device identity
Who enrolls, rotates, revokes, and retires device identities?
Managed identity helps only if it fits the device lifecycle and gateway model.
Enrollment test, revocation test, role map, and support owner.
Protocol and message shape
Are telemetry, command, files, retained state, and gateway bridges supported cleanly?
Custom brokers may need IaaS or containers; common ingestion can fit PaaS.
Protocol matrix, adapter test, retry behavior, and payload validation logs.
Data obligations
Who controls schema, retention, deletion, export, backup, restore, and audit access?
SaaS can accelerate workflow but may narrow export and retention choices.
Retention record, export rehearsal, restore evidence, and data owner signoff.
Operations capacity
Can the team patch, scale, monitor, recover, and support the owned layers?
IaaS increases control and operations burden; PaaS can reduce platform work.
Runbook, alert route, recovery drill, quota check, and escalation path.

Use the same worksheet for every boundary. A managed device registry may reduce certificate-service maintenance, but it does not decide whether a failed revocation blocks command delivery. A SaaS dashboard may reduce frontend work, but it does not prove raw export completeness, role separation, or support handoff. A self-managed broker may support unusual bridge behavior, but it also creates patch, capacity, backup, and on-call duties that need named owners.

Use IaaS When Control Is The Reason

Choose it when custom network isolation, broker behavior, runtime versions, protocol gateways, or deployment constraints are real requirements and the team can operate them.

Use PaaS When Managed Capability Fits

Choose it when managed identity, ingestion, queueing, data stores, functions, or streams meet the workload needs without hiding ownership of application and data behavior.

Use SaaS When Workflow Fit Is Strong

Choose it when a finished dashboard, alerting, work-order, or fleet-management surface fits the process and export, role, data, and integration boundaries are acceptable.

Common failure: treating “managed” as “no operations.” Managed services still require configuration review, quotas, identity policy, monitoring, incident response, data lifecycle checks, and migration planning.

3.3 Runtime Shape and Review Records

The service model and the compute pattern are separate decisions. A PaaS-oriented architecture can still use containers, functions, managed jobs, stream processors, or local gateway control. An IaaS architecture can still delegate some capabilities to managed services. The important detail is whether each runtime shape fits the state, timing, retry, observability, and ownership behavior of the workload.

IoT compute pattern route mapping event bursts, steady streams, stateful processors, scheduled jobs, and local control paths to functions, containers, services, managed jobs, and gateway control.
Compute choice is a behavioral decision inside the service-model boundary.

A short enrichment step can run well as AWS Lambda, Azure Functions, or a managed job when it is idempotent, observable, and safe to retry. A protocol bridge that holds MQTT sessions, translates payloads, and handles back-pressure may fit a container service such as Cloud Run, Azure Container Apps, ECS, or Kubernetes because it needs a controlled runtime and release path. A local alarm loop should stay in gateway firmware or an edge service when the WAN link cannot be part of the safety deadline. Those runtime choices can sit under IaaS, PaaS, or hybrid boundaries; the service model says who operates the layers, while the compute pattern says how the workload behaves.

The review record should tie both choices together. It should name the workload, service model, runtime shape, state owner, retry rule, identity boundary, telemetry store, recovery path, quota risk, release owner, and change signal. For example, a “PaaS ingestion” note is not enough for a freezer alert stream. The record needs the MQTT topic contract, certificate rotation owner, queue retention, late-message handling, dashboard freshness rule, export rehearsal, and gateway behavior when cloud acknowledgements stop.

This is also where teams catch mismatched promises: a serverless retry policy cannot fix a non-idempotent command, and a managed broker cannot prove local alarm behavior during a WAN outage.

Event Functions

Fit short independent reactions, enrichment, notification, and scheduled checks. Validate cold start, retry, idempotency, concurrency, and observability.

Containers And Workers

Fit sustained processing, custom runtimes, protocol bridges, stream consumers, and workloads that need controlled release and capacity behavior.

Long-Running Services

Fit persistent connections, stateful coordination, custom brokers, and low-latency loops. Keep restart, failover, and back-pressure evidence.

Local Gateway Control

Fit safety or availability paths that must continue during cloud outage. Define replay, duplicate detection, and reconciliation behavior.

IoT backend evolution path with four stages: one workload, module boundaries, separated workers, and independent services; a review gate checks ownership, scaling, release cadence, failure isolation, data boundaries, the 11-field review record, and the change signal before each split.
Split services when ownership, scaling, release cadence, failure isolation, or data boundaries require a new boundary and the 11-field review record supports the change.

Every service-model choice should leave a record. The record names the workload, the selected model, owned layers, delegated layers, compute pattern, data obligations, operations proof, portability assumptions, residual risks, owner, and change signal. Without that record, the architecture can drift into hidden obligations: unpatched virtual machines, unmonitored queues, unclear exports, brittle adapters, or a SaaS dependency with no exit path.

Cloud service model review record with workload intent, model choice, owned layers, delegated layers, compute pattern, data obligations, operations proof, portability, and owner signoff.
The review record is the handoff between architecture intent and operations evidence.

3.4 Summary

Cloud service models for IoT are ownership boundaries. IaaS gives more infrastructure control and more operating responsibility. PaaS can reduce platform work while leaving device, data, application, access, and operations decisions with the team. SaaS can accelerate dashboards, alerts, workflow, and fleet-management surfaces when configuration, data quality, integration, role design, export, and exit boundaries are acceptable. The right model is the one supported by workload behavior, operating capacity, security needs, data obligations, recovery evidence, and a clear review record.

3.5 Key Takeaway

Service-model choice is not just an implementation shortcut. It is an ownership decision that must name who operates each IoT layer, how the workload runs, what proof exists, and when the decision should be revisited.

3.6 See Also