Compute Engine is a customizable compute service that lets you create and run virtual machines on Google’s infrastructure. You can create a Virtual Machine (VM) that fits your needs. Predefined machine types are pre-built and ready-to-go configurations of VMs with specific amounts of vCPU and memory to start running apps quickly. With Custom Machine Types, you can create virtual machines with the optimal amount of CPU and memory for your workloads. This allows you to tailor your infrastructure to your workload. If requirements change, using the stop/start feature you can move your workload to a smaller or larger Custom Machine Type instance, or to a predefined configuration.

Machine types

In Compute Engine, machine types are grouped and curated by families for different workloads. You can choose from general-purpose, memory-optimized, compute-optimized and accelerator-optimized families. 

• General-purpose machines are used for Day-to-day computing at a lower cost and for balanced price/performance across a wide range of VM shapes. The use cases that best fit here are web serving, app serving, back office applications, databases, cache, media-streaming, microservices, virtual desktops, development environments.
• Memory-Optimized machine are recommended for ultra high-memory workloads such as in-memory analytics and large in-memory databases such as SAP HANA 
• Compute-Optimized machines are recommended for ultra high performance workloads such as High Performance Computing (HPC), Electronic Design Automation (EDA), gaming, video transcoding, single-threaded applications.
• Accelerator-Optimized machines are optimized for high performance computing workloads such as Machine learning (ML), Massive parallelized computations and High Performance Computing (HPC)

How does it work?

You can create a VM instance using a boot disk image, a boot disk snapshot, or a container image. The image can be a public operating system (OS) image or a custom one. Depending on where your users are you can define the zone you want the virtual machine to be created in. By default all traffic from the internet is blocked by the firewall and you can enable the HTTP(s) traffic if needed. 

Use snapshot schedules (hourly, daily, or weekly) as a best practice to back up your Compute Engine workloads. Compute Engine offers live migration by default to keep your virtual machine instances running even when software or hardware update occurs. Your running instances are migrated to another host in the same zone instead of requiring your VMs to be rebooted. 

Availability

For High Availability (HA) Compute Engine offers automatic failover to other regions or zones in event of a failure. Managed instance groups (MIGs) help keep the instances running by automatically replicating instances from a predefined image. They also provide application based autohealing health checks. If an application is not responding on a VM, the auto healer automatically recreates that VM for you. Regional MIGs let you spread app load across multiple zones. This replication protects against zonal failures. MIGs work with load balancing services to distribute traffic across all of the instances in the group. 

Compute Engine offers autoscaling to automatically add or remove VM instances from a managed instance group based on increases or decreases in load. Autoscaling lets your apps gracefully handle increases in traffic, and it reduces cost when the need for resources is lower. You define the autoscaling policy for automatic scaling based on the measured load, CPU utilization, requests per second or other metrics.

Active Assist’s new feature, predictive autoscaling, helps improve response times for your applications–When you enable predictive autoscaling, Compute Engine forecasts future load based on your Managed Instance Group’s (MIG) history and scales it out in advance of predicted load, so that new instances are ready to serve when the load arrives. Without predictive autoscaling, an autoscaler can only scale a group reactively, based on observed changes in load in real time. With predictive autoscaling enabled, the autoscaler works with real-time data as well as with historical data to cover both the current and forecasted load. That makes predictive autoscaling ideal for those apps with long initialization times and whose workloads vary predictably with daily or weekly cycles. For more information, see How predictive autoscaling works or check if predictive autoscaling is suitable for your workload, and to learn more about other intelligent features, check out Active Assist.

Pricing

You pay for what you use. But you can save cost by taking advantage of some discounts! Sustained use saving are automatic discounts applied for running instances for a significant portion of the month. If you know your usage upfront, you can take advantage of committed use discounts which can lead up to significant savings without any upfront cost. And by using short lived preemptive instances you can save up to 80%, they are great for batch jobs and fault tolerant workloads. You can also optimize resource utilization with automatic recommendations. For example if you are using a bigger instance for a workload that can run on a smaller instance you can save costs applying these recommendations.

Security

Compute Engine provides you default hardware security. Using Identity and Access Management (IAM) you just have to ensure that proper permissions are given to control access to your VM resources. All the other basic security principles apply, if the resources are not related and don’t require network communication amongst themselves, consider hosting them on different VPC networks. By default, users in a project can create persistent disks or copy images using any of the public images or any images that project members can access through IAM roles. You may want to restrict your project members so that they can create boot disks only from images that contain approved software that meet your policy or security requirements. You can define an organization policy that only allows Compute Engine VMs to be created from approved images. This can be done by using the Trusted Images Policy to enforce images that can be used in your organization. 

By default all VM families are Shielded VMs. Shielded VMs are virtual machine instances that are hardened with a set of easily configurable security features to ensure that when your VM boots, it’s running a verified bootloader and kernel — is the default for everyone using Compute Engine, at no additional charge. For more details on Shielded VMs refer to the documentation here.

For additional security, you also have the option to use Confidential VM to encrypt your data in use, while it’s being processed in Compute Engine. For more details on Confidential VM refer to the documentation here.

Use cases

There are many use cases Compute Engine can serve in addition to running websites and databases. You can also migrate your existing systems onto Google Cloud, with Migrate for Compute Engine, enabling you to run stateful workloads in the cloud within minutes rather than days or weeks. Windows, Oracle or VMware applications have solution sets enabling a smooth transition to Google Cloud. To run windows applications either bring your own license leveraging Sole-tenant nodes or using the included licenced images. 

Conclusion

Whatever your application use case may be, from legacy enterprise applications to digital native applications, Compute Engine’s families will fit it. For a more in-depth look into Compute Engine check out the documentation. 

For more #GCPSketchnote, follow the GitHub repo. For similar cloud content follow me on Twitter @pvergadia and keep an eye out on thecloudgirl.dev

Imagine ordering a “ready-to-assemble” table online, only to find that the delivery package did not include the assembly instructions. You know what the end product looks like, but have little to no clue how to start assembling the individual pieces to get there. A poorly designed API tends to create a similar experience for a consumer developer. Well designed APIs make it easy for consumer developers to find, explore, access, and use them. In some cases, good quality APIs even spark new ideas and open up new use cases for consumer developers.

There are methods to improve API design — like following RESTful practices. But time and again we are seeing customers unknowingly program minor inconveniences into their APIs. To help you avoid these pitfalls, here are six of the most common mistakes we have seen developers make while creating the API — and guidance on how to get it right.

#1 Thinking inside-out vs outside-in

Being everything for everybody often means that nothing you do is the best it could be, and that is just as true for APIs. When customers turn to APIs, they are looking for specific solutions to make their work easier and more productive. If there is an API that better works to their needs, they will choose that one over yours. This is why it’s so important to know what your customers need to do their work better, and then building to fill those needs. In other words, start thinking Outside-in as opposed to Inside-Out. Specifically,

• Inside-out refers to designing APIs around internal systems or services you would like to expose.
• Outside-in refers to designing APIs around customer experiences you want to create. Read more about the Outside-in perspective in the API product mindset.

The first step to this is learning from your customers — be it internal consumer developers or external customers — and their use cases. Ask them about the apps they are building, their pain points, and what would help streamline or simplify their development. Write down their most significant use cases and create a sample API response that only gives them the exact data they need for each case. As you test this, look for overlap between payloads and adapt your designs to genericize them across common or similar use cases.

If you can’t connect with your customers — because you don’t have direct access, they don’t have time, or they just don’t know what they want — the best approach is to imagine what you would build with your APIs. Think big and think creatively. While you don’t want to design your APIs for vaporware, thinking about the big picture can make it easier to build non-breaking changes in the future. For example the image below showcases APIs offered by Google Maps. Even without diving into the documentation, looking at the names like “Autocomplete” or “Address Validation” clearly outlines the purposes and potential fit for a customer’s use case.

#2 Making your APIs too complex for users

Customers turn to APIs to bypass complicated programming challenges so they can get to the part they know how to do well. If they feel like using your API means learning a whole new system or language, then it isn’t fitting their needs and they will likely look for something else. It’s up to your team to make an API that is strong and smart enough to do what your customer wants, but also simple enough to hide how complicated the tasks your API solves for really are. For example if you know your customers are using your APIs to present information about recently open restaurants and highly rated pizzeria to their consumers, providing them with a simple API call as below would be of great help:

GET /restaurants?location=Austin&category=Pizzeria&open=true&sort=-priority,created_at

To see if your API design is simple enough, pretend you are building the whole system from scratch — or if you have a trusted customer who is willing to help, ask them to test it and report their results. If you can complete the workflow without having to stop to figure something out, then you’re good to go. On the other hand, if you catch rough edges caused by trying to code around system complexity issues, then keep trying to refactor. The API will be ready when you can say that nothing is confusing and that it either meets your customers’ needs or can easily be updated as needs change.

#3 Creating “chatty” APIs with too many calls

Multiple network calls slow down the process and creates higher connection overhead — which means higher operational costs. This is why it’s so important to minimize the number of API calls.

The key to this is outside-in design: simplify. Look for ways to reduce the number of API calls a customer must make in their application’s workflow. If your customers are building mobile applications, for example, they often need to minimize their network traffic to reduce battery drain, and requiring a couple calls instead of a dozen can make a big difference.

Rather than deciding between building distinct, data-driven microservices and streamlining API usage, consider offering both: fine-grained APIs for specific data types, and “experience APIs” (APIs that are designed to power user experiences. Here is a further theoretical discussion on Experience APIs) around common or customer-specific user interfaces. These experience APIs compose multiple smaller domains into a single endpoint; making it much simpler for your customers — especially those building user interfaces — to render their screens easily and quickly.

Another option here is to use something like GraphQL to allow for this type of customizability. Generally you should avoid building a unique endpoint for every possible screen, but common screens like home pages and user account information can make a world of difference to your API consumers.

#4 Not allowing for flexibility

Even if you’ve followed all of the steps above, you may find that there are edge cases that do not fit under your beautifully designed payloads. Maybe your customer needs more data in a single page of results than usual, or the payload has way more data than their app requires. You can’t create a one-size-fits-all solution, but you also don’t want a reputation for building APIs that are limiting. Here are 3 simple options to make your endpoints more flexible.

Filter out response properties: You can either use query parameters for sorting and pagination, or use GraphQL which provides these types of details natively. By giving customers the option to request only the properties they need, it guarantees that they won’t have to sort through tons of unnecessary data to get what they need. For example, if some of your customers only need the title, author, and bestseller ranking, give them the ability to retrieve only that data with a query string parameter.

GET /books?fields=title,author,ranking

• Ability to sort with pagination. Generally, you don’t want to guarantee the order of objects in an API response because minor changes in logic or peculiarities in your data source might change the sort order at some point. In some cases, however, your customers may want to sort by a particular field. Giving them that option, combined with a pagination option, will give them a highly efficient API when they only want the top few results. For example Spotify API utilizes a simple offset and limit parameter set to allow pagination. A sample endpoint as shown in the documentation would look like this

$ curl https://api.spotify.com/v1/artists/1vCWHaC5f2uS3yhpwWbIA6/albums?album_type=SINGLE&offset=20&limit=10

• Use mature compositions like GraphQL: Since customer data needs can differ, giving them on-the-fly composites lets them build to the combinations of data they need, rather than being restricted to a single data type or a pre-set combination of data fields. Using GraphQL can even bypass the need to build experience APIs, but when this isn’t an option, you can use query string parameter options like “expand” to create these more complex queries. Here is a sample response that demonstrates a collection of company resources with embedded properties included

"data": [
{
"CompanyUid": "27e9cf71-fca4",
"name": "ABCCo",
"status": "Active",
"_embedded": {
"organization": {
"CompanyUid": "27e9cf71-fca4",
"name": "ABCCo",
"type": "Company",
"taxId": "0123",
"city": "Portland",
"notes": ""
}
}
}
]

#5 Making design unreadable to humans

“K”eep “I”t “S”imply “S”tupid when you are designing your API. While APIs are meant for computer-to-computer interaction, the first client of an API is always a human, and the API contract is the first piece of documentation. Developers are more apt to study your payload design before they dig into your docs. Observation studies suggest that developers spend more than 51% of their time in editor and client as compared to ~18% on reference.

For example, if you skim through the payload below it takes some time to understand because instead of property names it includes an “id”. Even the property name “data” does not suggest anything meaningful aside from just being an artifact of the JSON design. A few extra bytes in the payload can save a lot of early confusion and accelerate adoption of your API. Notice how user-ids appearing on the left of the colon (in the position where other examples of JSON ideally have property names) creates confusion in reading the payload.

"{id-a}":
{ "data":
[
{
"AirportCode": "LAX",
"AirportName": "Los Angeles",
"From": "LAX",
"To": "Austin",
"departure": "2014-07-15T15:11:25+0000",
"arrival": "2014-07-15T16:31:25+0000"
}
… // More data
]
},

We think that JSON like this is more difficult to learn. If you want to eliminate any ambiguity in the words you choose to describe the data, keep the payload simple and if any of those labels could be interpreted in more than one way, adjust them to be more clear. Here is a sample response from Airlines endpoint of aviationstack API. Notice how the property names clearly explain the expected result while maintaining a simple JSON structure.

"data": [
{
"airline_name": "American Airlines",
"iata_code": "AA",
"iata_prefix_accounting": "1",
"icao_code": "AAL",
"callsign": "AMERICAN",
"type": "scheduled",
"status": "active",
"fleet_size": "963",
"fleet_average_age": "10.9",
"date_founded": "1934",
"hub_code": "DFW",
"country_name": "United States",
"country_iso2": "US"
},
[…]
]

#6 Know when you can break the RESTful rules

Being true to the RESTful basics — such as using the correct HTTP verbs, status codes, and stateless resource-based interfaces — can make your customers’ lives easier because they don’t need to learn an all new lexicon, but remember that the goal is just to help them get their job done. If you put RESTful design first over user experience, then it doesn’t really serve its purpose.

Your goal should be helping your customers be successful with your data, as quickly and easily as possible. Occasionally, that may mean breaking some “rules” of REST to offer simpler and more elegant interfaces. Just be consistent in your design choices across all of your APIs, and be very clear in your documentation about anything that might be peculiar or nonstandard.

Conclusion

Beyond these common pitfalls, we have also created a comprehensive guide packaging up our rich experience designing and managing APIs at incredible scale with Google Cloud’s API management product, Apigee.

Apigee — Google Cloud’s native API management platform — helps you build, manage, and secure APIs — for any use case, scale or environment. Get started with Apigee today or check out our documentation for additional information.

Introduction

In my previous post, I talked about how you can use a parent workflow to execute child workflows in parallel for faster overall processing time and easier detection of errors. Another useful pattern is to use a Cloud Tasks queue to create Workflows executions and that’s the topic of this post.

When your application experiences a sudden surge of traffic, it’s natural to want to handle the increased load by creating a high number of concurrent workflow executions. However, Google Cloud’s Workflows enforces quotas to prevent abuse and ensure fair resource allocation. These quotas limit the maximum number of concurrent workflow executions per region, per project, for example, Workflows currently enforces a maximum of 2000 concurrent executions by default. Once this limit is reached, any new executions beyond the quota will fail with an HTTP 429 error.

A Cloud Tasks queue can help. Rather than creating Workflow executions directly, you can add Workflows execution tasks to the Cloud Tasks queue and let Cloud Tasks drain the queue at a rate that you define. This allows for better utilization of your workflow quota and ensures the smooth execution of workflows.

Let’s dive into how to set this up. 

Create a Cloud Tasks queue

We’ll start by creating a Cloud Tasks queue. The Cloud Tasks queue acts as a buffer between the parent workflow and the child workflows, allowing us to regulate the rate of executions.

Create the Cloud Tasks queue (initially with no dispatch rate limits) with the desired name and location:

QUEUE=queue-workflow-child
LOCATION=us-central1

gcloud tasks queues create $QUEUE --location=$LOCATION

Now that we have our queue in place, let’s proceed to set up the child workflow.

Create and deploy a child workflow

The child workflow performs a specific task and returns a result to the parent workflow.

Create workflow-child.yaml to define the child workflow:

main:

  params: [args]

  steps:

    - init:

        assign:

          - iteration: ${args.iteration}

    - wait:

        call: sys.sleep

        args:

            seconds: 10

    - return_message:

        return: ${"Hello world" + iteration}

In this example, the child workflow receives an iteration argument from the parent workflow, simulates work by waiting for 10 seconds, and returns a string as the result.

Deploy the child workflow:

gcloud workflows deploy workflow-child --source=workflow-child.yaml --location=$LOCATION

Create and deploy a parent workflow

Next, create a parent workflow in workflow-parent.yaml.

The workflow assigns some constants first. Note that it’s referring to the child workflow and the queue name between the parent and child workflows:

main:
  steps:
    - init:
        assign:
          - project_id: ${sys.get_env("GOOGLE_CLOUD_PROJECT_ID")}
          - project_number: ${sys.get_env("GOOGLE_CLOUD_PROJECT_NUMBER")}
          - location: ${sys.get_env("GOOGLE_CLOUD_LOCATION")}
          - workflow_child_name: "workflow-child"
          - queue_name: "queue-workflow-child"

In the next step, Workflows creates and adds a high number of tasks (whose body is an HTTP request to execute the child workflow) to the Cloud Tasks queue:

- enqueue_tasks_to_execute_child_workflow:
        for:
          value: iteration
          range: [1, 100]
          steps:
              - iterate:
                  assign:
                    - data:
                        iteration: ${iteration}
                    - exec:
                        # Need to wrap into argument for Workflows args.
                        argument: ${json.encode_to_string(data)}
              - create_task_to_execute_child_workflow:
                  call: googleapis.cloudtasks.v2.projects.locations.queues.tasks.create
                  args:
                      parent: ${"projects/" + project_id + "/locations/" + location + "/queues/" + queue_name}
                      body:
                        task:
                          httpRequest:
                            body: ${base64.encode(json.encode(exec))}
                            url: ${"https://workflowexecutions.googleapis.com/v1/projects/" + project_id + "/locations/" + location + "/workflows/" + workflow_child_name + "/executions"}
                            oauthToken:
                              serviceAccountEmail: ${project_number + "-compute@developer.gserviceaccount.com"}

Note that task creation is a non-blocking call in Workflows. Cloud Tasks takes care of running those tasks to execute child workflows asynchronously.

Deploy the parent workflow:

gcloud workflows deploy workflow-parent --source=workflow-parent.yaml --location=$LOCATION

Execute the parent workflow with no dispatch rate limits

Time to execute the parent workflow:

gcloud workflows run workflow-parent --location=$LOCATION

As the parent workflow is running, you can see parallel executions of the child workflow, all executed roughly around the same:

In this case, 100 executions is a well under the concurrency limit for Workflows. Quota issues may arise if you submit 1000s of executions all at once. This is when Cloud Tasks queue and its rate limits become useful.

Execute the parent workflow with dispatch rate limits

Let’s now apply a rate limit to the Cloud Tasks queue. In this case, 1 dispatch per second:

gcloud tasks queues update $QUEUE --max-dispatches-per-second=1 --location=$LOCATION

Execute the parent workflow again:

gcloud workflows run workflow-parent --location=$LOCATION

This time, you see a more smooth execution rate (1 execution request per second):

Summary

By introducing a Cloud Tasks queue before executing a workflow and playing with different dispatch rates and concurrency settings, you can better utilize your Workflows quota and stay below the limits without triggering unnecessary quota related failures. 

Check out the Buffer HTTP requests with Cloud Tasks codelab, if you want to get more hands-on experience with Cloud Tasks. As always, feel free to contact me on Twitter @meteatamel for any questions or feedback.