Skip to main content
DevOps NetworkDevOpsNetwork
Join Free

Reducing Compute Costs with Spot VMs and Custom Machine Types

Learn to cut real Compute Engine spend using Spot VMs for fault-tolerant batch work and custom machine types sized to actual workload needs.

38 Terms

Overview and What You Will Learn

In this lab, you will create a Spot VM for a batch workload and observe its discounted pricing and interruption behavior, then create a custom machine type sized to a workload's actual needs rather than the nearest predefined size.

Why This Matters in Production

A team runs a nightly batch processing job on a standard predefined machine type that's noticeably larger than the job actually needs, simply because it was the closest predefined size available. A custom machine type sized to the job's real CPU and memory requirements, combined with Spot pricing for this genuinely interruption-tolerant workload, meaningfully reduces the cost of a job that runs every single night indefinitely.

Core Principles

Spot VMs and custom machine types solve two different cost problems and should not be confused with each other.

◈ DIAGRAM
+------------------------------------------+
| Spot VMs |
| For: fault-tolerant batch or stateless |
| workloads that can handle interruption |
| Effect: steep discount, but Google can |
| reclaim the VM with short notice |
+------------------------------------------+
| Custom Machine Types |
| For: any workload whose actual needs fall |
| between two predefined sizes |
| Effect: pay for exactly the vCPU/memory |
| combination needed, no more |
+------------------------------------------+

Spot VMs are appropriate specifically for workloads that can tolerate being interrupted and restarted elsewhere - never for anything customer-facing that needs to be reliably available. Custom machine types are appropriate whenever a workload's real requirements don't cleanly match a predefined size, regardless of whether that workload is fault-tolerant or not.

Detailed Step-by-Step Practical Lab

  1. Create a project and enable Compute Engine:
Bash
gcloud projects create gcp-cost-lab-2026 --name="Compute Cost Lab"
gcloud config set project gcp-cost-lab-2026
gcloud services enable compute.googleapis.com
  1. Create a Spot VM for a batch-processing workload:
Bash
gcloud compute instances create vm-batch-spot \
--zone=asia-south1-a \
--machine-type=e2-standard-4 \
--image-family=debian-12 \
--image-project=debian-cloud \
--provisioning-model=SPOT \
--instance-termination-action=STOP
Note

--instance-termination-action=STOP means an evicted Spot VM is stopped rather than deleted, preserving its disk and configuration so it can be restarted later - similar in spirit to Azure's DeleteOnTermination-equivalent choice or AWS Spot's stop-vs-terminate interruption behavior.

  1. Confirm the VM was created with the expected Spot pricing model:
Bash
gcloud compute instances describe vm-batch-spot \
--zone=asia-south1-a \
--format="value(scheduling.provisioningModel)"
  1. Create a custom machine type sized to a specific, real workload need - here, 6 vCPUs and 20GB memory, a combination that falls between two standard predefined sizes:
Bash
gcloud compute instances create vm-custom-shape \
--zone=asia-south1-a \
--custom-cpu=6 \
--custom-memory=20GB \
--image-family=debian-12 \
--image-project=debian-cloud
  1. Compare the custom machine type's actual specification against the nearest predefined alternatives:
Bash
gcloud compute instances describe vm-custom-shape \
--zone=asia-south1-a \
--format="value(machineType)"
gcloud compute machine-types list \
--filter="zone:asia-south1-a AND name~'e2-standard'" \
--format="table(name,guestCpus,memoryMb)"
  1. Check current Recommender suggestions for any existing VMs that might be over-provisioned:
Bash
gcloud recommender recommendations list \
--project=gcp-cost-lab-2026 \
--location=asia-south1-a \
--recommender=google.compute.instance.MachineTypeRecommender
  1. Clean up:
Bash
gcloud compute instances delete vm-batch-spot vm-custom-shape \
--zone=asia-south1-a --quiet
gcloud projects delete gcp-cost-lab-2026 --quiet

Production Best Practices & Common Pitfalls

Common Mistake

Using a Spot VM for a customer-facing or otherwise critical workload to save money, then experiencing an unexpected outage when Google reclaims the capacity with only a short notice period. Spot pricing is a genuine discount, but it comes with a real availability trade-off that only fits fault-tolerant, interruption-tolerant workloads.

Tip

Check Recommender's machine type suggestions periodically for existing VMs - it analyzes actual historical CPU and memory utilization and suggests a better-fitting predefined or custom machine type, often surfacing meaningful savings on VMs that were sized as a rough guess at launch time and never revisited.

  • Custom machine types are billed based on the specific vCPU and memory configuration chosen, not rounded up to the nearest predefined size - this is what makes them genuinely cost-effective for a workload whose real needs fall between two standard offerings.
  • Combining Spot pricing with a custom machine type stacks both savings for a workload that is both fault-tolerant and has non-standard resource needs - a batch job with unusual memory requirements is a strong candidate for both optimizations simultaneously.

Quick Reference & Troubleshooting Commands

Command Description
gcloud compute instances create --provisioning-model=SPOT Create a Spot VM
gcloud compute instances create --custom-cpu= --custom-memory= Create a custom machine type VM
gcloud compute machine-types list List available predefined machine types
gcloud recommender recommendations list Check machine type rightsizing suggestions

Resources

Artifact Registry

Google Cloud's current, unified package and container image storage service, the successor to the older Container Registry. It supports container images alongside language packages like npm, Maven, and Python packages in one service, and supports cleanup policies to automatically remove old, untagged image versions.

Defined Term Details

BigQuery

Google Cloud's serverless data warehouse for analytics, priced based on the volume of data a query actually scans rather than pre-provisioned compute capacity. Selecting only needed columns and querying a properly partitioned table's relevant date range are the two highest-impact levers for controlling BigQuery cost.

Defined Term Details

Bigtable (GCP)

A wide-column NoSQL database built for very high throughput at extremely low latency, suited to time-series data, IoT telemetry, and analytics workloads with massive write volume. It is a specialized tool distinct from Firestore's document model or Cloud SQL's relational model.

Defined Term Details

Billing Account (GCP)

A separate object from a Project that holds the actual payment method for Google Cloud usage. One Billing Account can be linked to many Projects, all of whose usage is billed to that same account, letting a company centralize payment while individual Projects retain their own separate IAM control over resources.

Defined Term Details

Cloud Armor

A security policy service attached to a Global External HTTP(S) Load Balancer, filtering malicious requests before they reach the backend. It supports rate-based rules blocking a single source IP exceeding a request threshold, and preconfigured WAF rules blocking common attack patterns like SQL injection, Google's equivalent of AWS WAF or Azure WAF.

Defined Term Details

Cloud Functions (GCP)

Google Cloud's event-driven, single-purpose compute service, the most granular and shortest-lived compute unit in GCP's compute spectrum. It fits a narrow niche compared to Cloud Run - a specific triggered action like resizing an uploaded image, rather than a full application with multiple routes and business logic.

Defined Term Details

Cloud Load Balancing

Google Cloud's family of load balancing products, spanning Global External HTTP(S) Load Balancers (public, distributed globally via a single IP), Regional External Load Balancers, and Internal Load Balancers reachable only from within the VPC. The right type depends on traffic type and whether the backend should ever be reachable from the public internet.

Defined Term Details

Cloud Logging

Google Cloud's service for collecting, storing, and querying log data from every GCP service and custom applications. Its query filter syntax lets logs be narrowed by resource, severity, and specific field values, and Log Sinks can route matching logs to BigQuery, Cloud Storage, or Pub/Sub for long-term retention beyond the default window.

Defined Term Details

Cloud Monitoring

Google Cloud's metrics, dashboards, and alerting service. An alert policy requires a Notification Channel explicitly attached to actually notify someone when it fires - without one, the policy still evaluates and logs its own firing history, but nobody is ever told, a common and easy-to-miss misconfiguration.

Defined Term Details

Cloud NAT

A service providing outbound-only internet connectivity for resources with no external IP address, GCP's equivalent of an AWS or Azure NAT Gateway. Cloud NAT requires a Cloud Router in the same region and provides zero inbound connectivity capability - a resource behind it remains completely unreachable from the internet.

Defined Term Details

Cloud Run

A fully managed, serverless platform for running containerized applications, scaling automatically to zero when idle and billing only for actual invocations. It combines the flexibility of running any container with zero OS management overhead, making it the right default for a standard stateless containerized web service.

Defined Term Details

Cloud SQL

Google Cloud's managed relational database service, supporting MySQL, PostgreSQL, and SQL Server. It is the right default for a traditional relational workload with moderate scale needs, handling patching, backups, and replication automatically without requiring OS-level access to the underlying database server.

Defined Term Details

Cloud Spanner

A globally distributed, strongly consistent relational database offering both SQL semantics and horizontal scalability beyond what a single Cloud SQL instance can provide. It is the right choice specifically when a workload needs both relational query capability and massive scale simultaneously, not simply because a workload feels large or important.

Defined Term Details

Cloud Storage Bucket

The top-level container for objects in Google Cloud Storage, GCP's object storage service equivalent to AWS S3 or Azure Blob Storage. Every object uploaded to Cloud Storage must belong to exactly one bucket, and bucket-level settings control default storage class, location, and access permissions for everything inside it.

Defined Term Details

Cloud VPN (GCP)

A service connecting an on-premises network to a GCP VPC over an encrypted tunnel, similar to VPN Gateway in Azure or a Site-to-Site VPN in AWS. It requires a Cloud Router as its foundation and is distinct from VPC Network Peering, which connects two GCP VPCs directly rather than an external network.

Defined Term Details

Compute Engine

Google Cloud's Infrastructure-as-a-Service offering, providing full virtual machines with complete OS-level control. It is the right choice specifically when a workload needs custom software, specific OS configurations, or hardware-level access that GKE and Cloud Run don't expose.

Defined Term Details

Error Reporting (GCP)

A Google Cloud operations service that automatically aggregates and groups application errors from logs and instrumented code, surfacing recurring error patterns and their frequency without requiring manual log searching for each individual occurrence.

Defined Term Details

Firestore

A serverless NoSQL document database built for mobile and web application data, flexible schemas, and real-time client synchronization. It is not a strong fit for workloads needing complex relational joins or multi-row transactions across unrelated entities, which point instead to Cloud SQL or Spanner.

Defined Term Details

GCP Audit Log

A specific category of Cloud Logging entry recording administrative actions and data access. Admin Activity logs (configuration changes) are always enabled and cannot be disabled; Data Access logs (who read or wrote actual data) are disabled by default for most services and must be explicitly enabled before that history exists.

Defined Term Details

GCP Firewall Rule

A rule defined at the VPC level controlling inbound or outbound traffic, applied using network tags or Service Accounts on instances rather than being tied to a specific subnet boundary. A firewall rule only affects VMs carrying the exact matching tag or identity - a VM without it is entirely unaffected, silently, with no error to flag the mismatch.

Defined Term Details

GCP Folder

An optional grouping node in the GCP resource hierarchy, sitting between the Organization and individual Projects. Folders are commonly used to group Projects by team, department, or environment, letting IAM permissions and Organization Policies be applied once at the Folder level rather than repeated across every Project inside it.

Defined Term Details

GCP IAM Role Types

Google Cloud IAM roles come in three types: Basic roles (Owner, Editor, Viewer) which are broad and legacy; Predefined roles curated by Google and scoped to a specific service's specific needs; and Custom roles built from individual permissions for access needs a predefined role doesn't precisely match.

Defined Term Details

GCP Organization

The top-level node in Google Cloud's resource hierarchy, tied to a Google Workspace or Cloud Identity domain. An Organization contains Folders and Projects, and policies or IAM permissions set at this level automatically apply to everything beneath it across the entire company.

Defined Term Details

GCP Project

The fundamental unit of organization in Google Cloud - every resource belongs to exactly one Project, which serves as the boundary for billing, IAM permissions, and API enablement. A Project has a permanent, globally unique Project ID chosen at creation and never changeable, distinct from its display name which can be updated freely.

Defined Term Details

GCP Quota

A default limit on how much of a specific resource (like VM CPUs or API calls) can be used per Project per region, designed to prevent runaway costs and protect shared infrastructure. Quotas can be checked and increased through the Console ahead of a known future need, such as a planned launch event.

Defined Term Details

GCP Storage Class

A pricing and access tier for Cloud Storage objects - Standard, Nearline, Coldline, or Archive - trading lower storage cost against higher retrieval cost and a longer minimum storage duration. Deleting or overwriting an object before its class's minimum duration elapses still bills as if it were stored the full minimum period.

Defined Term Details

GKE Autopilot

A mode of Google Kubernetes Engine where Google manages node provisioning, sizing, and patching entirely - you define workloads, and nodes appear automatically to run them, billed per pod resource request rather than per node. It is the lower-effort default for most containerized workloads, versus Standard mode's direct node pool control.

Defined Term Details

Managed Instance Group

A group of identical Compute Engine VMs created from an Instance Template and managed as a single unit, GCP's equivalent of an AWS Auto Scaling Group or Azure VM Scale Set. A Managed Instance Group can autoscale based on a metric like CPU utilization, adding or removing instances automatically within configured minimum and maximum bounds.

Defined Term Details

Network Tag (GCP)

A freeform string label attached to a Compute Engine VM at creation time, used to target firewall rules to specific VMs. Network tags are simple but can be forgotten or misspelled at scale, which is why Service Account-based targeting is often preferred for anything security-critical in a growing fleet.

Defined Term Details

OS Login

A Compute Engine feature that ties SSH access to a user's actual Google Identity and IAM permissions, rather than manually managing individual SSH public keys per VM. Enabling OS Login means revoking a departing employee's GCP access also immediately revokes their SSH access to every VM, with no separate key cleanup needed.

Defined Term Details

Organization Policy (GCP)

A constraint applied at the Organization, Folder, or Project level that restricts what resources are allowed to look like across everything beneath that level - for example, blocking external IP addresses on VMs or restricting allowed regions. Unlike IAM, which controls who can act, Organization Policy controls what configurations are permitted.

Defined Term Details

Persistent Disk (GCP)

The block storage attached to a Compute Engine VM, equivalent to an AWS EBS volume or Azure Managed Disk. Persistent Disks come in zonal (replicated within one zone, lower cost) and regional (synchronously replicated across two zones, survives a single zone failure) variants.

Defined Term Details

Security Command Center

Google Cloud's security posture management service, continuously scanning deployed resources against security best practices and surfacing specific, severity-ranked findings - like a publicly accessible storage bucket - rather than a single opaque score. Findings should be prioritized by actual severity, not treated as a count to minimize.

Defined Term Details

Service Account (GCP)

An identity used by applications, VMs, and automated processes to authenticate to Google Cloud APIs, rather than by a human user. A Service Account is itself also a resource with its own IAM permissions, and can be attached directly to a VM or impersonated temporarily, both safer alternatives to downloading its long-lived JSON key file.

Defined Term Details

Shared VPC

A configuration where one Project (the host) owns a VPC network, while other Projects (service projects) attach to it and deploy resources using its subnets. This centralizes network administration in the host Project while letting individual teams in service Projects still manage their own resources independently.

Defined Term Details

Spot VM (GCP)

A Compute Engine VM using Google's spare, unused capacity at a steep discount, with the trade-off that Google can reclaim the instance with short notice. Spot VMs are appropriate only for fault-tolerant, interruption-tolerant workloads like batch processing, never for customer-facing services needing reliable availability.

Defined Term Details

VPC (GCP)

A Virtual Private Cloud in Google Cloud is a global resource by default, unlike the regional VPC or VNet model in AWS or Azure - a single GCP VPC can span every region, with individual subnets defined per-region inside it. The default auto-mode VPC pre-creates a subnet in every region, while a custom-mode VPC starts empty until subnets are deliberately defined.

Defined Term Details

VPC Network Peering

A direct connection between two separate VPCs, potentially in different Projects or Organizations, with traffic never touching the public internet. Peering must be established from both sides to become active, and is explicitly non-transitive - if VPC A peers with B, and B peers with C, A cannot reach C without its own direct peering connection to C.

Defined Term Details

Explore More in GCP Compute and Container Services

All 6 Topics

Choosing Between Compute Engine, GKE, Cloud Run, and Cloud Functions

Learn how to choose the right GCP compute service based on control needs, container orchestration complexity, and traffic patterns.

Explore Topic

Configuring Managed Instance Groups with Autoscaling

Learn to build a Managed Instance Group from an Instance Template with autoscaling based on real CPU load, GCP's equivalent of an ASG.

Explore Topic

Deploying to GKE Autopilot vs Standard Mode

Learn the real trade-offs between GKE Autopilot's fully managed nodes and Standard mode's direct node pool control.

Explore Topic

Deploying Serverless Containers with Cloud Run

Learn to deploy a container to Cloud Run, configure traffic splitting for gradual rollouts, and set scaling parameters correctly.

Explore Topic

Frequently Asked Questions

Is Reducing Compute Costs with Spot VMs and Custom Machine Types free to learn on DevOps Network?

Yes - this topic, like everything on DevOps Network, is 100% free with no paywall or sign-up gate.

What does the Reducing Compute Costs with Spot VMs and Custom Machine Types topic cover?

Learn to cut real Compute Engine spend using Spot VMs for fault-tolerant batch work and custom machine types sized to actual workload needs.

Back to GCP Hub