Prepare and test your skills
Prepare and test your skills
Workload modernization and modification is the strategy of refactoring or re-platforming an existing application to replace self-managed infrastructure with Google Cloud managed services. The goal is to reduce operational work for patching, backups, availability, and scaling while improving scalability and resilience. Modification makes sense when storage, compute, and configuration dependencies have a compatible managed replacement and when lower management effort outweighs migration and retraining costs.
Modernized compute moves work from manually managed virtual machines to managed groups, containers, or serverless platforms that handle availability and scaling. Compute Engine managed instance groups (MIGs) (opens in a new tab) automatically recreate failed virtual machines (VMs) and, for stateless workloads, can autoscale with traffic, while regional MIGs stay resilient to zone outages. Google Kubernetes Engine (GKE) (opens in a new tab) automates deploying, scaling, and managing containerized applications, and regional GKE clusters are also resilient to zone outages. Cloud Run is a serverless platform where code runs on a preconfigured platform with minimal management effort.
The choice between compute platforms trades configuration flexibility for management effort:
| Compute option | Configuration flexibility | Management effort |
|---|---|---|
| Compute Engine VMs with MIGs | High, full control of VM and operating system | High, team manages VMs, scaling, and recovery |
| Google Kubernetes Engine (GKE) | High, control of containers and orchestration | Medium, Google manages control plane, team manages workloads |
| Cloud Run | Low, deploy to preconfigured platform | Low, minimal infrastructure operation |
Choose VMs or GKE when the application needs operating system access, custom networking, or container orchestration. Choose Cloud Run when the priority is to focus on data and applications instead of setting up and operating infrastructure.
Modernized data management replaces self-managed databases and disks on VMs with managed services that provide high availability backed by service level agreements (SLAs), built-in backups, and scalability. Cloud SQL provides high availability configurations, read replicas across zones, and included database license costs, while a third-party database on Compute Engine VMs forces the team to apply updates, monitor availability, perform backups, and recover from failures. Spanner is a distributed, multi-regional database that provides built-in redundancy and synchronization across regions for active-active topologies. Cloud Storage (opens in a new tab) provides regional, dual-region, and multi-region buckets for durable object storage without managing disks.
Storage modernization also widens locational scope to improve resilience. Regional persistent disks provide synchronous replication of data across two zones within a region, which improves availability over zonal disks. Filestore regional instances replicate data synchronously across three zones within a region and can be mounted on multiple VMs for shared configuration files, tools, and logs. For dual-region Cloud Storage buckets, turbo replication ensures objects are replicated across the region pair with a recovery point objective (RPO) of 15 minutes.
A modern data platform separates the data storage layer from the data computation layer while keeping them colocated in the same zone and region. The data storage layer is where data is saved, either as files on a file system or through a database management system. The data computation layer loads data from the storage layer, processes the data, and then saves results to a target system, which can be the source storage layer.
Data flows from the storage platform to the computation platform and then to the target system, so traffic flows from storage to compute and back to storage. To reduce communication latency, outbound data transfer cost, and input/output operations, the data is stored in the same zone that processes the data. To improve flexibility and reduce resource use, storage and computation run on separate platforms in the same zone and region, so the computation layer does not have to run all the time. For example, data storage can move from Hadoop Distributed File System (HDFS) to Cloud Storage while computation runs on a Managed Service for Apache Spark cluster. Common storage building blocks are Cloud Storage, BigQuery storage, Cloud SQL, Spanner, and Bigtable, while common computation building blocks are Dataflow, Managed Service for Apache Spark, BigQuery, custom workloads on GKE or Compute Engine, and notebooks.
Exam tip: Keep storage and computation on separate platforms for flexibility, but in the same zone and region to reduce latency and transfer cost.
Assessment determines requirements, dependencies, and costs before refactoring or re-platforming batch pipelines and applications. The assessment follows an ordered sequence:
Technical feasibility depends on compatibility of each layer. For each storage platform, the team verifies that the product is available in the target environment or has a compatible replacement, and practices the data transfer process because each platform has its own migration method. For each computation platform, the team inventories custom configurations and versions, such as Apache Spark or Apache Kafka versions, because different versions can change outputs, serializations, and computation. The team tests network latency between the source and target environments and from clients to the target, and notes cluster sizes such as node count, CPUs, and memory, because a move to another region can change the processor.
Business impact depends on scalability, reliability, and cost controls. The team benchmarks resource efficiency, which is the relationship between application load and CPU and memory consumed, under both normal and peak conditions and repeats tests regularly because efficiency can deteriorate under high load. The design plans N+1 redundancy so every component has at least one redundant component beyond the minimum needed for forecast load, and overprovisions by a third when using three zones per region to handle failover. The team reviews Google Cloud quotas, which are per-project limits that can be zonal, regional, or global, including resource quotas for maximum resources and rate quotas for API requests, and requests an adjustment when quotas are insufficient for planned capacity.
A build decision means creating a custom solution on platforms such as Compute Engine or Google Kubernetes Engine (GKE), including custom virtual machines, custom operating system images, and self-managed databases. A buy decision means using managed commercial services such as Cloud SQL (opens in a new tab), Spanner, Cloud Storage, BigQuery, and Dataflow, which include built-in backup, high availability, and scalability. The strategic trade-off centers on competitive advantage, total cost of ownership (TCO), speed to market, and who carries operational responsibility for reliability and maintenance.
Custom builds give maximum control over configuration, version, and architecture, but the team owns the underlying compute, networking, and storage infrastructure. Examples include running Oracle Database on Compute Engine VMs (opens in a new tab), building custom Windows bring-your-own-license (BYOL) images, maintaining custom image families, and running custom workloads on GKE or Compute Engine. This path can support differentiation when no commercial product fits, but it is slower to launch and requires designing availability, replication, failover, and observability from scratch.
Managed services shift much of that work to Google Cloud, which provides high availability backed by uptime service-level agreements (SLAs), built-in backup mechanisms, and resilience features. Regional managed instance groups (MIGs) and GKE regional clusters are resilient to zone outages, while managed multi-regional services such as Spanner and Cloud Storage improve resilience against region outages. When designing either path, there is a trade-off between availability, operational complexity, latency, and cost, because increasing the number of nines of availability generally costs more. For example, 99.9% availability allows about 86 seconds of downtime in 24 hours, while 99% allows 864 seconds, which is 10 times more downtime.
Total cost of ownership for a target environment includes resource sizing, licensing, storage performance, and ongoing operations, not just hourly VM cost. Correct sizing is critical because disk performance depends on disk size and number of virtual CPUs (vCPUs), and incorrect sizing can cause long migration times, database errors, and application performance problems. Key sizing inputs are disk size and throughput, input/output operations per second (IOPS), and number of vCPUs, and costs can be estimated with the Google Cloud Pricing Calculator.
Financial assessment is supported by dedicated tools and cost controls:
Licensing changes TCO between paths. With Cloud SQL, database license costs are included in the service charge, while Oracle products on Compute Engine require the customer to procure licenses and comply with Oracle terms. For Microsoft Windows Server, a custom Windows BYOL image means paying only for Google Cloud infrastructure, which preserves value from existing licenses. Storage choice also matters: solid-state drive (SSD) is the most efficient choice for most uses, while hard disk drive (HDD) may fit datasets of 10 terabytes or more that are latency-insensitive or infrequently accessed.
Operational ownership is the largest difference between build and buy. Managed services handle synchronous replication between zones within a region, automatic recreation of failed VMs through MIGs, load balancing across resources, and high availability configurations for databases. A custom build must implement each control, including distributing resources to eliminate a single point of failure (SPOF), planning N+1 redundancy with at least one redundant component beyond minimum load, reserving Compute Engine zonal capacity, managing resource and rate quotas, and automating provisioning and recovery to reduce human error.
Aggregate availability shows why operational design matters. The aggregate availability of chained tiers is tier1_availability × tier2_availability × tierN_availability, so four tiers at 99.9% each yield only about 99.6% overall. Chained services at 99% each yield about 98%, while the same two services in parallel yield about 99.99% using 1 − (1 − A) × (1 − B) because each service runs independently. Buying managed multi-regional storage, using regional persistent disks with synchronous replication, distributing traffic with Cloud DNS and regional or global load balancers, and separating the data storage layer from the data computation layer in the same zone and region improve both reliability and flexibility.
Exam tip: High availability (HA) generally applies to regional deployments with synchronous replication and zero or near-zero recovery time and recovery point objectives, while disaster recovery (DR) generally applies to multi-region deployments with typically asynchronous replication and defined failover operations.
A structured assessment connects business requirements to the build-versus-buy choice by inventorying current systems, testing feasibility, and calculating TCO before committing. The assessment phase for batch and database workloads follows an ordered sequence:
This inventory work exposes where buying is faster and cheaper and where building is justified. If assessment shows heavy proprietary features, strict data sovereignty limits on failover regions, or unique performance needs such as C4 machine series VMs and zonal internal DNS for Oracle connectivity, a custom build may be required. If managed equivalents exist, support open export formats such as CSV, Parquet, and Avro, and meet availability needs, buying accelerates speed to market while regulators still require exit plans for retrieving data and maintaining continuity if a provider relationship ends.
Workload deprecation and retirement is the decision to remove a system, database, table, or infrastructure stack that no longer justifies its cost, risk, or operational load. A retirement decision balances declining business value and technical debt against security and compliance needs, then preserves required data and shifts dependent services before anything is deleted. In Google Cloud, this work connects Cloud SQL (opens in a new tab) inventory and monitoring signals to archival in Cloud Storage (opens in a new tab) and to an ordered teardown of dependent networking and compute resources.
Candidate identification starts from usage, cost, and reliability signals that show a workload no longer needs its current footprint. Cloud Monitoring provides predefined and custom dashboards for health and capacity of components including Cloud SQL, while Active Assist provides cost insights and recommendations for idle and overprovisioned resources. A workload becomes a retirement or rightsizing candidate when monitoring shows persistent low use, excess provisioning, or components that do not require high availability or disaster recovery.
The retirement signals include:
Retirement is not only deletion. The same assessment may lead to changing storage class, consolidating instances, purchasing committed use discounts for predictable workloads, or keeping only automated backups for less-critical data.
Secure archival preserves the final state of a retiring workload for compliance, audit, and possible restore. Archival in Cloud Storage moves data from the live database or disk to a full backup or an archival Cloud Storage bucket before the source is deleted. Because regulators may require proof of the end state, a final backup of each source database is taken before cleanup and retained under retention controls.
Archival controls separate what must be kept from what can expire, and they protect retained data from early deletion. Retention policies with Bucket Lock, temporary holds on active objects, versioning, uniform bucket-level access, blocked public access, logging of key bucket activities, and lifecycle and deletion rules together define how long objects are kept and when they can be removed. Exporting to open formats such as CSV, Parquet, and Avro, and preferring services based on open technologies, preserves the ability to retrieve and move data if a provider relationship ends.
Exam tip: Create a final backup of each source database before deleting it, and keep at least 20% available disk space on a Cloud SQL instance with a metrics-based alert for maintenance operations.
Dependent-service protection ensures that applications, replicas, and operations continue while a legacy system is retired. An exit plan defines strategies and contingency plans for data retrieval and operational continuity if a service or provider relationship ends, including assessment of exit feasibility without operational disruption. The plan names the systems on both sides of each dependency, so traffic, replication, and maintenance do not break when one side is removed.
The transition moves in order: first validate the target, then cut over, then observe, then clean up. After cutover, applications that use the target database are monitored for a defined period to decide whether a fallback plan is needed, with awareness that rolling back may cost more effort than fixing forward. For database cutovers, source parameter settings and statistics are collected and compared with the target, a maintenance window is configured on the primary instance, and writes on the target can be replicated back to the original source so the old system stays current until retirement is confirmed. An administrative operation is started only after the previous operation has completed.
Ordered teardown deletes a retired stack from the edge inward so no deleted resource is still referenced by another resource. The load-balanced stack depends in one direction: proxies and certificates depend on the URL map, the URL map depends on the backend service, the backend service depends on health checks and managed instance groups (MIGs), MIGs depend on the instance template, and the template depends on the image and disk. Deletion therefore flows from frontend to backend to template to data.
A top-to-bottom process flow showing the nine-step deletion order for a retired load-balanced stack, from target proxies down to the source disk.
The documented removal sequence is:
The same principle applies to databases: delete the source database only after the final backup is secured, parameters and statistics are captured, dependent applications are stable on the target, and any required reverse replication for fallback is no longer needed.
Choose Cloud Run when the priority is to focus on data and applications instead of setting up and operating infrastructure, because it runs code on a preconfigured serverless platform with minimal management effort. Choose Compute Engine VMs with MIGs or GKE when the application needs operating system access, custom networking, or container orchestration, accepting higher management effort for full control of VMs and operating systems or control of containers and orchestration.
High availability generally applies to regional deployments with synchronous replication and zero or near-zero recovery time and recovery point objectives. Disaster recovery generally applies to multi-region deployments with typically asynchronous replication and defined failover operations.
Delete the retired stack from frontend to backend in dependency order, starting with the HTTPS and HTTP target proxies, then the SSL certificate, URL map, backend service, and health check, followed by the zonal managed instance groups, instance template, custom image, and finally the source disk. Each resource is removed only after the resources that depend on it are gone, so the certificate is removed after no proxy references it and the template is removed after the MIGs that depend on it are gone.
An online retail enterprise is modernizing its enterprise applications and planning its Google Cloud migration. The architecture review board is evaluating disposition strategies for two distinct systems:
The organization has finite software engineering resources and must optimize Total Cost of Ownership (TCO), minimize ongoing operational burden, and maximize competitive advantage.
Which workload disposition strategy should the cloud architect recommend?
Last updated