Cloud foundations / Cloud Engineering
Cloud engineering.
Foundations, workloads and recovery.
Plan the accounts, identity, network, data and operating model beneath your application. Scope provider-specific infrastructure, reviewed changes, migration and recovery around the workload and the team that receives it.
Discuss Cloud Engineering requirements βContents Β· Cloud Engineering
Build the foundation before choosing the workload platform.
Cloud engineering can cover account or subscription structure, identity, network addressing, compute, data services, delivery, monitoring, cost and recovery. Start with existing systems, private-data flows, governance and receiving-team needs rather than installing a cluster by default.
Plan Terraform modules, remote state and reviewed changes before workloads. Inspect and test reusable infrastructure against the actual accounts, regions, data and operating responsibilities. A module library still needs workload-specific configuration, access and recovery checks.
Modules are provider-specific. A consistent review workflow across AWS, Azure and Google Cloud does not make one provider resource definition execute unchanged on another. Define portability for application, data, identity and operation separately.
Account control, configuration rights and service terms differ.
Agree who owns organization/account/subscription/project administration, billing, recovery contacts, domains, keys, identity and state. Define named roles and revocation rather than assuming a contractor-held account can be transferred later without provider constraints.
Handover should include application and infrastructure definitions, modules, versions, charts, workflow files, architecture and network maps, state location, backup/restore steps, dashboards, alerts, certificate renewal and receiver exercises. References to secrets belong in documentation; private values do not.
Custom code assignment, reusable bases, third-party modules, packages, images and managed-service terms require separate review. You control agreed configuration and data in managed services; you do not acquire the providerβs implementation or an unconditional right to export every managed feature.
Source delivery does not guarantee no lock-in or immediate independent operation. Verify permissions, dependencies, data export, tool compatibility and the receiverβs ability to change, release and recover the system.
Compare operating responsibilities rather than supplier labels.
- An internal team can own architecture, review, incident response and product context when its skills, capacity and coverage fit. Hiring or employment alone does not establish those controls.
- A project team can deliver defined foundations, migration and receiver handoff. Verify actual people, review evidence, account permissions, acceptance and the boundary after delivery.
- A managed service provider can operate agreed services and support coverage. Check customization, tool/data portability, access, escalation, pricing and exit terms in the specific contract.
- For each model compare architecture authority, change process, security evidence, availability coverage, cost, documentation and recovery. A company is not automatically accountable for every outcome, and an individual is not inherently a console-only operator.
Do not adopt the original hiring delays, ticket response intervals, proprietary-tool stereotypes or guaranteed delivery comparison. Choose the actual arrangement that meets the workload and receiving-team needs.
Seven questions define the cloud engagement.
What is in scope?
Name accounts, networks, identity, compute, data, delivery, observability, cost, migration and runbooks. The application itself and ongoing response can be separate deliverables.
Cloud, DevOps or Kubernetes?
Cloud engineering covers foundations and platform services; DevOps covers the delivery/operating path; container work covers packaging and orchestration. Agree interfaces where they overlap.
Which provider?
Evaluate required services, regions, existing identity/licensing, skills, private connectivity, governance, cost and recovery rather than a universal ranking.
Is multi-cloud needed?
A requirement may justify another provider, but it adds separate identity, network, billing, service semantics and operating dependencies. An exit plan can also be tested without running every workload active-active.
How does migration run?
Inventory dependencies, choose rehost/replatform/refactor per workload, build the target, rehearse data movement and authorize staged cutover with post-write recovery.
Who pays and controls usage?
Separate implementation and ongoing fees from provider/tool bills. Agree account roles, limits, estimates, tagging and cost-response ownership before spend.
What uptime and recovery are needed?
Define workload availability, RTO/RPO, data invariants and failure scenarios. A managed-service SLA or multi-zone diagram is not proof of end-to-end recovery.
Separate foundation, workload and observation decisions.
Foundation and infrastructure definitions
Choose account boundaries, non-colliding address ranges, private connectivity, SSO, roles, network policies and reviewed provider resources. Inventory unmanaged exceptions and bootstrap dependencies rather than promising every resource is already code.
Workload placement and delivery
Compare existing hosts, managed containers, Kubernetes and functions against runtime, state, traffic, identity, availability and receiver needs. If a cluster is selected, define its capacity, network, access, release and upgrade owners.
Observability and operation
Use metrics, logs, traces, release events, certificate/backup checks and cost signals with recipients and runbooks. Test representative performance and failure paths; neither a development laptop nor an unrepresentative staging stack proves production behaviour.
The original says staging is never used as a performance benchmark while its QA section prescribes production-like staging tests. Resolve that into a representative-test requirement with measured differences, not an unsupported prohibition or claim a real load test already happened.
The root module references infrastructure it does not define.
The example prod/main.tf sets a Terraform lower version bound, an S3 backend with encryption and native locking, an AWS provider constraint/default tags, and local network/EKS-cluster modules. It is an illustrative root configuration, not a complete account foundation.
- The S3 bucket must already exist. encrypt=true and use_lockfile=true do not create the bucket, enable its versioning, define IAM/KMS policy or prove state backup and recovery. Backend bootstrap needs separate authority and ownership.
- Native S3 locking is enabled in this sample, but current backend defaults do not enable it automatically. The state and lockfile need their own permitted object actions. DynamoDB-based locking is deprecated in current documentation.
- The backend fixes eu-west-1 while the provider uses var.region. Changing the provider region does not automatically move the state backend or create a new environment. Review keys, buckets, roles and account boundaries before init or migration.
- required_version>=1.11 is a lower bound, not an exact tested version or a major-version ceiling. The AWS ~>6.0 constraint permits compatible6.x selections, not one frozen build. Review and commit the dependency lock file for actual provider selections/checksums.
- The provider lock file does not lock remote module versions. The local ../modules/network and ../modules/eks-cluster code must be supplied and versioned with the repository. Its inputs are custom module interfaces, not proof of their implementation.
- az_count=3 does not show subnet distribution, route tables, capacity or a tested zone failure. node_min=3/node_max=10 are module inputs; they do not establish actual node autoscaling, pod HPA, ready replicas or unlimited capacity.
- public_endpoint=false is also a custom module input. Verify how it maps to EKS endpoint settings, private DNS, runner/operator network access and IAM/RBAC. A private Kubernetes API does not itself define public application access or complete security.
- default_tags does not guarantee every resource and child resource is tagged. AWS provider tagging support has exceptions, including Auto Scaling groups requiring explicit treatment, and cost-allocation activation/reporting remains separate.
- Variables, credentials, provider permissions, network/EKS module contents, CI plan/apply, identity, secrets, policy, monitoring, backups and load/restore tests are not supplied. A file called prod does not establish an isolated production account.
- A successful plan/apply would concern selected infrastructure, not restore application data or prove the whole system is reproducible in another provider. Inspect the actual account, region, resource replacements and live result.
Treat the code below as a partial root configuration. Inspect the module implementations, state, IAM, network, workload and cost before applying a reviewed plan in an authorized account.
Technical example
Source sample, not an installed backend or verified deployment. The limitations above apply.
# prod/main.tf - one environment, its own root module and state
# Variables (region, owner, cost_center) are declared in variables.tf
terraform {
required_version = ">= 1.11"
backend "s3" {
bucket = "acme-platform-tfstate"
key = "prod/platform.tfstate"
region = "eu-west-1"
encrypt = true
use_lockfile = true # S3-native state locking
}
required_providers {
aws = {
source = "hashicorp/aws"
version = "~> 6.0"
}
}
}
provider "aws" {
region = var.region
# Default tags cover supported resources; verify exceptions and billing activation
default_tags {
tags = {
environment = "prod"
owner = var.owner
cost_center = var.cost_center
managed_by = "terraform"
}
}
}
module "network" {
source = "../modules/network"
cidr = "10.20.0.0/16"
az_count = 3 # spread across three availability zones
}
module "cluster" {
source = "../modules/eks-cluster"
subnet_ids = module.network.private_subnet_ids
public_endpoint = false
node_min = 3
node_max = 10
}Protect state, locks and the reviewed change artifact.
Choose backend access, encryption, versioning, locking, retention and recovery deliberately. Separate operational state from application backups. A state snapshot can reveal resource identifiers and secrets; marking a value sensitive can hide display output without removing it from state or plans.
Bind production apply to the intended reviewed plan and destination under the chosen process. Re-running plan at apply time can propose different changes. Saved plans themselves can contain sensitive data and need protected storage, access and retention.
Define concurrent writer prevention, interrupted-apply reconciliation and lock recovery. Do not force-unlock a writer because an automation job appears old. Provider errors and partially changed resources require current-state inspection.
Import can associate existing resources with Terraform, but requires correct identity, address, provider and compatible configuration. Review the resulting plan for unwanted replacement, deletion or field changes. Importing does not guarantee no production effect or complete management of every setting.
Landing-zone guardrails need concrete scope.
Define organizations or management groups/folders, production/non-production separation, SSO, federation, break-glass recovery, workload roles and approval rights within existing governance. Account separation does not remove application tenant-access obligations.
AWS service control policies set permission limits; they do not grant IAM access and do not constrain the organization management account in the same way as member accounts. Review exact attachment scope, exceptions and effective permissions before treating a policy label as protection.
Azure and Google Cloud governance need their own product-specific controls and evidence. A Terraform workflow does not make all policy semantics identical. Prove the selected configuration rather than claiming an AWS example establishes all-cloud enforcement.
Grant only required access for the permitted work. Record activity, credential lifetime and revocation. A bilateral NDA or source-code promise does not supply authorization to access private production data or change billing.
Addressing and private access are prerequisites, not decoration.
Inventory current networks, remote users, partners and expected VPN/private links before assigning address ranges. Document routing, DNS, subnet roles, egress, security groups/firewalls and endpoint dependencies. A subnet called private is not proof no path exists.
Private API endpoints and private workload/data endpoints can require different routes, identity and DNS. Make CI and authorized operators able to reach the intended endpoint without granting broad public access as a workaround.
Inspect Internet/CDN/load-balancer and service-to-service paths separately. Network segmentation does not replace application authentication, authorization, tenant policy or protected administrative access.
Name certificates, renewal, trust, connection draining and traffic-shift responsibilities. DNS or route changes require their own approved destination, window and recovery plan. Confirm the actual account, destination and recovery route before a change.
Choose a platform without turning a pattern into a mandate.
Managed containers, functions, VMs and Kubernetes expose different runtime, scaling, identity, network and operations choices. A standard module can help selected workloads; it does not prove one platform fits every application or a cluster is needed for a small service.
Map relational/document/key-value records, cache/queue state and object storage to actual consistency, latency, throughput, retention and restore requirements. Managed service names do not establish a backup policy or correct business transactions.
Separate application scaling from node or provider capacity. Quotas, addresses, service limits, downstream databases, hot partitions, startup and recovery can constrain capacity even when an autoscaling policy exists.
Define upgrade and compatibility review for runtimes, providers, images, controllers, APIs and managed services. Source delivery is incomplete if the receiving team cannot identify supported versions and a safe change path.
Cloud controls support compliance work, not certification claims.
Review who can read customer data, secrets, logs, artifacts, state and backups. Keep plaintext credentials out of source and ordinary output; secret references still need scoped access, rotation and recovery.
A selected workload region does not prove all data stays there. Inspect CI, global identity/CDN/control-plane services, logs, replicas, backups, support access and subprocessors. Region-deny policies can contain global-service exceptions.
Encryption, least privilege, network segmentation, logging and policy checks support a control programme. They do not certify SOC2, PCI DSS, HIPAA or GDPR compliance, replace required agreements or prove every organizational obligation is met.
Managed services change the provider/customer boundary, not eliminate customer responsibilities. Verify the actual service, features, location and contract. Define permitted records, secret storage and diagnostic retention for the actual deployment.
Availability design and recoverability need different evidence.
A multi-zone architecture can reduce selected zone-failure risks, but is not a second-region plan or a backup. Shared dependencies, identity, DNS, application bugs and data corruption can still affect all replicas.
Define recovery time and recovery point objectives for meaningful workload functions and data. Test detection, decision, access, keys, restore/replay, configuration, traffic and user-facing acceptance, not only whether a VM boots.
Replicas can reproduce deletes or corruption. Protect appropriate independent backup history and test restore with the receiving team. Confirm the database, object, volume, queue and application semantics rather than assuming one cloud snapshot covers everything.
Provider SLAs are service-specific contracts with conditions and remedies. Combining services does not automatically inherit a blanket application uptime guarantee. Do not publish measured RTO/RPO or failover success without permitted test evidence.
Rehearse the data boundary before traffic moves.
Inventory servers, databases, scheduled jobs, dependencies, identity, certificates, external callbacks and hidden manual steps. Choose rehost, replatform, refactor or retention per workload rather than claiming one method is always fastest or cheapest.
Build and validate the target foundation with representative workloads. Review replication/log retention, conversion, lag, validation, private-data access and a bounded change freeze where needed. Continuous replication is not a guarantee the final sync is immediate or loss-free.
Stage traffic only after health, business invariants and data readiness checks. Lower TTLs do not force every cache or client to stop using the old address. A routing plan is not a transaction-coordination protocol.
After new writes, the old system can be stale. Agree reverse replay, reconciliation, forward repair or freeze according to the workload. Retain old capacity, backups, keys and access until accepted exit criteria, then decommission under authority.
Desired state needs owners and recovery rules.
CI can build, test, scan and promote reviewed artifacts; GitOps can reconcile selected desired resources. Define identity, branch/environment controls, runner privileges, plan/artifact binding and the actual production mutation owner.
Auto-sync, prune, self-heal and ignored fields are chosen controls, not automatic proof drift cannot occur. Avoid multiple controllers or scripts managing the same field with conflicting intentions.
A revert or infrastructure rollback does not undo database migrations, published messages, payments or every external effect. Test version compatibility and forward/reverse recovery before presenting deployment history as a complete safety net.
Verify live artifact/configuration, representative business paths and observability after release. A green job or successful apply is progress, not full application acceptance.
Monitoring that runs continuously is not continuous staffing.
Define service health, latency, error, saturation, queue/dependency and business-invariant signals. Use metrics, structured logs, traces and release context with appropriate data minimization, access and retention.
Test alerts, certificate/backup failures, budget notifications and escalation with actual recipients. A dashboard, CloudWatch rule or Slack integration does not itself respond to an incident.
Specify staffed coverage, response expectations, authority, break-glass access and communication in current terms. The source distinguishes monitoring around the clock from engineering working hours; do not convert that into an always-staffed help desk or adopt a stale response interval.
Receiver runbooks should cover diagnosis, scaling, release recovery, restore, credentials, certificate renewal and provider escalation. Exercise the steps rather than delivering only screenshots and tool names.
Three models have distinct acceptance and support boundaries.
Architecture and foundation
Scope provider/account structure, network, identity, guardrails, IaC and the reviewed-change pipeline. Require diagrams, definitions, current estimates and receiver evidence rather than assuming a generic base contains everything.
Migration or custom cloud work
Define application/data movement, provider-specific architecture, integrations, cutover, post-write recovery and acceptance. Some scopes require application engineering alongside platform work.
Managed infrastructure
Agree patching, cost/capacity/security reviews, new environments, incident coverage, change approval and exit. Agree one current written cancellation and support policy with explicit notice, scope and exit responsibilities.
Six gates make the proposed standard testable.
- Review:architecture, account/region and plan changes with approved resource effects and exceptions.
- Automated checks:format/validate, lint, policy and appropriate application/infrastructure tests; not proof every issue is excluded.
- Representative load and failure:production-like configuration and documented differences, scaling/failover/restore exercises with approved blast radius.
- Handoff:definitions, state/access references, versions, network/identity maps, dashboards, runbooks and receiver operation.
- Cutover:data readiness, traffic, DNS/certificate checks, stop conditions and post-write recovery.
- Operation:tested alerts, backups, cost signals, security maintenance and named response owners.
A destructive zone or database exercise needs scoped permission and safeguards. Agree the actual drill, stop conditions, recovery evidence and incident/support responsibilities before touching an environment.
Sixteen stack roles are selected for the actual platform.
- AWS:chosen compute, storage, network, data and identity services.
- Azure:subscription, Microsoft identity, network, hybrid and workload services.
- Google Cloud:project/folder governance, data, compute and selected managed services.
- Terraform:provider-specific desired resources, plans and protected state.
- Kubernetes:orchestration when the workload and receiver justify it.
- Docker:reviewed image/runtime inputs and architecture compatibility.
- Ansible:scoped configuration changes with appropriate access.
- Helm:rendered workload definitions, values and version-aware recovery.
- Prometheus:metrics and selected alert rules.
- Grafana:dashboards and signal access.
- CloudWatch:AWS-specific logs, metrics and alarm paths.
- Datadog:chosen telemetry/integration scope and retention terms.
- GitLab CI:reviewed workflows, runners, identity and artifacts.
- GitHub Actions:versioned workflows, runner isolation and environment controls.
- Vault:selected secrets/identity integration, policies and recovery.
- Consul:selected discovery/mesh capabilities with their own operation.
Argo CD/Flux, provider monitoring, OpenTofu, managed databases, caches, queues and storage may also be scoped. A list is not proof installed features, current licenses, staffing or universal inclusion.
Five checkpoints link scope to a usable handoff.
- Brief:understand existing application, data, governance, private access and authorized participants.
- Plan:provider/service model, account roles, versions, acceptance, current quote and recovery route.
- Build:review definitions and representative environments with real evidence and explicit manual dependencies.
- Validate:test workload/security, load, failover, restore, alerts and cost assumptions proportionate to scope.
- Handoff:approve any actual cutover, verify receiver access and runbooks, and agree ongoing responsibilities.
The original first-commit, NDA, architecture, custom-build, catalogue and support intervals are not current guarantees. Obtain milestones and dependencies in writing; account approvals, provider quotas, data migration and receiver readiness can affect the plan.
Cost control is an operating process, not a cap label.
Separate application catalogue work, architecture/foundation, migration/custom work and managed operation. Accounts, providers, regions, environments, workloads, data volume, network, availability, private connectivity and required controls affect the quote.
Estimate compute, data services, storage/request/retention, network/NAT/egress, CI, telemetry, backup and human operation. Right-sizing and retention changes must preserve accepted reliability and legal requirements; an idle-looking resource can still be a dependency.
Tags help selected attribution but not every resource supports the same tags or propagates them to child resources. Verify billing activation, untagged/shared charges and report scope. A provider default_tags block is not a complete cost ledger.
AWS Budgets notifications can lag usage/billing and do not universally stop spend. Budget actions have supported-resource/permission and approval boundaries. Other provider controls need their own review. Commitments and paid changes require separate owner approval before enabling resources or commitments.
Get one current itemized quote and support/cancellation terms. Use current written terms for the selected scope, estimate and support.
Proxy-product scope and cloud responsibilities to define.
A supplied proxy/VPN product narrative describes a custom proxy/VPN application and says no product base fitted. It also records17 of21 building blocks reused unchanged from a custom-build base,4 custom pieces,3 applications/consoles across development/staging/production and6 isolated integrations.
Those are recorded specification counts, not independently established delivery, reuse, IP transfer or production outcomes. Ask what the reusable base contained, what was custom, which components/integrations were used and what permitted acceptance evidence exists. Do not silently describe all work as from scratch.
The stack is only described as chosen for the project. Do not invent AWS, Azure, Google Cloud, Kubernetes, Terraform or a particular database implementation from that phrase. A recorded delivery year is not evidence of current service operation.
The separate review cards cover Maxx VPN orchestration/subscriptions, Server ProGuard/MXSecurity Dashboard alerts/collectors and PAS Systems live-video features. They do not independently prove a cloud migration or measured availability, cost or performance improvement.
Supplier diligence should ask for evidence, not repeat claims.
Ask for permitted technical examples, real demo limits, account control, source/dependency rights, reference consent, acceptance/restore evidence and current support/exit terms. Verify the actual contracting identity separately.
Do not adopt ratings, awards, project/client counts, launch/store approval, staffing, cost/speed or reliability outcomes. Corporate-registration assertions remain excluded. Facts, leadership and review links are leads for verification, not proof of every claim.
Apply the same questions to project teams, individuals, internal operations and managed providers. Geography or an organization label does not determine engineering quality, security or accountability.
Eight domains change the platform requirements.
- Healthcare/life sciences:private records, processing locations, access, audit, retention and recovery.
- Telemedicine:consultation video/messaging, privacy, availability and outage handling.
- Fintech:ledger invariants, provider callbacks, settlement, reconciliation and regulatory responsibility.
- Retail/e-commerce:stock, orders, payments and peak-demand behaviour.
- Media/entertainment:upload/transcoding queues, rights, storage and delivery.
- Creator economy:feeds, uploads, subscriptions, moderation and payouts.
- Transportation/mobility:location freshness, dispatch, connectivity and failure modes.
- Food/beverage:outlets, menus, order transitions and demand peaks.
A domain card establishes neither a certified client nor a guarantee every record remains in-region, no downtime occurs or all traffic peaks are absorbed. Verify application, platform and organizational controls together.
Questions about cloud engineering.
Which provider fits?
Evaluate actual services, regions, identity, skills, governance, network, cost and recovery rather than claiming one provider is universally widest, best or cheapest.
What will cloud engineering cost?
Obtain current architecture/migration/custom/managed scope and separate provider/tool estimates. Record assumptions, exclusions, milestones and support in the proposal.
What does ownership mean?
Define account/billing control, source rights, state, keys, data export and receiver operation. Managed services retain provider terms; source delivery is not zero lock-in proof.
How long does migration take?
Inventory dependencies, data, conversion, approvals and acceptance, then agree waves and milestones. No universal elapsed-time or no-downtime promise applies.
Cloud, on-premises or hybrid?
Compare actual latency, residency, connectivity, legacy dependencies, skills, reliability and total operating cost. Cloud is not automatically the lower-cost or safer choice.
What belongs in managed support?
Name monitoring, staffed response, patching, cost/security/capacity reviews, restore tests, change authority and exit. Automated monitoring is not continuous staffing.
Does cloud engineering certify compliance?
No. Product controls support agreed requirements; organizational obligations, required agreements, audits and attestations need their own evidence.
What happens after deployment?
Verify acceptance, receiver access, runbooks, backups, signals and response ownership. Obtain current defect/support and ongoing-change terms.
How are confidentiality and rights handled?
Agree contracting identity, NDA, permitted private-data access, custom/base/third-party rights and revocation.
Can application and cloud work be combined?
Yes when separately scoped with features, data, platform, delivery and receiver acceptance. A catalogue application does not automatically include a complete cloud foundation.
Can existing enterprise governance be retained?
Work within approved account/subscription/folder structure, identity, roles, naming, tagging and review controls. Verify exact policies and permitted changes.
Can manually built infrastructure move into code?
Audit resources and state, associate supported resources with correct definitions, and review plans for mutation. Import is not universal coverage or a guarantee production never changes.
Primary references for technical qualification.
Official platform, framework and API documentation checked on 17 November 2026. These support the technical checks here, not Omnistu staffing, pricing, rights, delivery or client outcomes. Recheck installed versions and configuration before implementation.
- Backend Type: s3 | Terraform | HashiCorp Developer - developer.hashicorp.com
- Dependency Lock File (.terraform.lock.hcl) - Configuration Language | Terraform | HashiCorp Developer - developer.hashicorp.com
- Configure default tags for AWS resources | Terraform | HashiCorp Developer - developer.hashicorp.com
- aws_ecs_service | Resources | hashicorp/aws | Terraform | Terraform Registry - registry.terraform.io
- aws_appautoscaling_target | Resources | hashicorp/aws | Terraform | Terraform Registry - registry.terraform.io
- Import resources overview | Terraform | HashiCorp Developer - developer.hashicorp.com
- Manage sensitive data in your configuration | Terraform | HashiCorp Developer - developer.hashicorp.com
- Availability Zones - AWS Fault Isolation Boundaries - docs.aws.amazon.com
- Service control policies (SCPs) - AWS Organizations - docs.aws.amazon.com
- Shared Responsibility Model - Amazon Web Services (AWS) - aws.amazon.com
- module block reference | Terraform | HashiCorp Developer - developer.hashicorp.com
- Version Constraints - Configuration Language | Terraform | HashiCorp Developer - developer.hashicorp.com
- lifecycle meta-argument reference | Terraform | HashiCorp Developer - developer.hashicorp.com
- Cluster API server endpoint - Amazon EKS - docs.aws.amazon.com
- Pass Secrets Manager secrets through Amazon ECS environment variables - Amazon Elastic Container Service - docs.aws.amazon.com
- Managing your costs with AWS Budgets - AWS Cost Management - docs.aws.amazon.com
- Terraform AWS Provider Resource Tagging | Guides | hashicorp/aws | Terraform | Terraform Registry - registry.terraform.io
Bring the domain, dependencies and receiving-team needs.
Describe the records, users, permitted actions, integrations, existing code and operating constraints. Agree base/custom boundaries, acceptance evidence, current quote and handover before implementation.
Discuss Cloud Engineering requirements β