PCDOE Optimizing service performance — Timed 25 Questions

Question 1

Your team has deployed a microservices application on Google Kubernetes Engine (GKE). You notice that one service has high latency during peak hours. The service is CPU-bound and uses a HorizontalPodAutoscaler (HPA) based on CPU utilization. What is the most likely cause of the latency?

25 minute time limit — choose an answer to begin.

Accepted Answer

The HPA's target CPU utilization is set too high, causing the autoscaler to react slowly.. When the HPA's target CPU utilization is set too high, the autoscaler waits until the average CPU utilization exceeds that threshold before scaling up. During peak hours, the service becomes CPU-bound and latency increases as pods are overwhelmed, but the HPA reacts slowly because it only triggers when the high threshold is breached, causing a delay in adding new pods to handle the load.

Answer

The GKE cluster uses preemptible nodes that are frequently reclaimed.

Answer

The service uses a global external HTTP(S) load balancer with session affinity.

Answer

The application does not implement request autoscaling at the application layer.

Question 2

You are designing a globally distributed application using Cloud Spanner. The application has a write-heavy workload. You notice that write latency increases as the number of nodes increases. What is the most likely cause?

Accepted Answer

The workload has many cross-node transactions due to split rows.. In Cloud Spanner, write-heavy workloads with many cross-node transactions cause increased write latency as nodes are added. This occurs because Spanner splits rows across nodes, and transactions that span multiple splits require two-phase commit (2PC) coordination between nodes, which adds network overhead and latency. Adding more nodes increases the likelihood that a transaction touches multiple splits, exacerbating the coordination cost.

Answer

The instance is using a multi-region configuration with too many read-only replicas.

Answer

The application is using stale reads for write transactions.

Answer

The number of splits is too low, causing hotspots.

Question 3

A company runs a stateful workload on Compute Engine VMs with persistent disks. They observe that disk I/O latency spikes periodically. The workload is sensitive to latency. What should they do to improve performance?

Accepted Answer

Use SSD persistent disks instead of standard persistent disks.. SSD persistent disks provide consistent, low-latency I/O performance compared to standard persistent disks, which use spinning media and can exhibit periodic latency spikes under sustained load. For latency-sensitive stateful workloads, SSD persistent disks offer predictable IOPS and throughput, directly addressing the periodic spikes observed.

Answer

Increase the size of the persistent disk.

Answer

Migrate to local SSDs for better performance.

Answer

Configure a snapshot schedule to offload I/O.

Question 4

Your GKE cluster runs a batch job that processes large files from Cloud Storage. The job uses CPUs inefficiently, with low utilization. You want to reduce cost while maintaining throughput. Which approach should you take?

Accepted Answer

Use Cloud Storage FUSE to stream files directly into containers, avoiding local storage.. Cloud Storage FUSE allows containers to stream files directly from Cloud Storage without first downloading them to a local disk. This eliminates the I/O bottleneck of writing to local storage and reduces CPU overhead from disk operations, enabling the batch job to process files more efficiently and maintain throughput while using fewer CPU resources.

Answer

Configure the node pool to use spot VMs.

Answer

Use local SSDs for faster file access.

Answer

Increase the CPU request for the job pods.

Question 5

You are using Cloud CDN with an external HTTPS load balancer. Users in Asia report slow load times for static assets. The origin is in us-central1. What should you do to improve performance?

Accepted Answer

Enable Cloud CDN and configure cache modes for static content.. Cloud CDN caches static assets at Google's global edge locations, reducing latency for users in Asia by serving content from a nearby point of presence instead of the us-central1 origin. Enabling Cloud CDN and configuring cache modes for static content (e.g., setting Cache-Control headers or using origin cache policies) ensures that frequently requested assets are served from cache, dramatically improving load times for geographically distant users.

Answer

Switch the load balancer to an internal HTTPS load balancer with gRPC.

Answer

Use premium tier networking for the load balancer.

Answer

Configure a serverless NEG to route traffic to Cloud Functions.

Question 6

You are using Memorystore for Redis as a cache for a high-traffic web application. You observe that cache hit ratio is low, causing high database load. What is the most effective way to improve cache hit ratio?

Accepted Answer

Use the allkeys-lru eviction policy to keep frequently accessed keys.. The allkeys-lru eviction policy is the most effective way to improve cache hit ratio because it automatically evicts the least recently used keys across the entire keyspace when memory is full, retaining the most frequently accessed data. This ensures that the cache always contains the hottest data, directly increasing the likelihood of cache hits without requiring manual intervention or additional resources.

Answer

Increase the instance size to store more data.

Answer

Migrate to Memorystore for Memcached for lower latency.

Answer

Set a longer TTL for all cache entries.

Question 7

Which THREE factors should you consider when designing a Cloud Run service for optimal performance under unpredictable traffic patterns? (Choose 3)

Accepted Answer

Set a minimum number of instances to reduce cold starts.. Option B is correct because setting a minimum number of instances keeps warm instances available, which directly reduces cold-start latency when traffic spikes unpredictably. Option D is correct because allocating sufficient CPU per instance ensures each instance can handle peak request load without throttling, since Cloud Run CPU limits affect request processing speed and concurrency capacity. Option E is correct because setting a maximum number of instances bounds scaling, controlling concurrency and cost so unpredictable traffic does not cause runaway scaling or unexpected billing. Option A is not marked correct because HTTP/2 connection reuse is a client-to-service optimization and is not a primary design factor for handling unpredictable traffic scaling in Cloud Run. Option C is not marked correct because Cloud NAT affects outbound VPC egress latency and IP availability, not the core performance tuning of a Cloud Run service under variable inbound traffic.

Answer

Use HTTP/2 for faster connection reuse.

Answer

Configure VPC egress through Cloud NAT for lower latency.

Question 8

Which TWO metrics from Cloud Monitoring would best indicate that a GKE workload is experiencing CPU throttling due to a resource quota? (Choose 2)

Accepted Answer

container/cpu/throttled_time. Option B, container/cpu/throttled_time, is correct because this metric directly measures the cumulative time a container's CPU usage was throttled by the Linux CFS (Completely Fair Scheduler) quota mechanism, which is the definitive signal of CPU throttling due to a resource quota (Kubernetes CPU limits). Option D, container/cpu/usage_time, is correct because it shows the actual CPU time consumed by the container, and when correlated with throttled_time it reveals whether the workload is hitting its CPU limit and being constrained, confirming throttling is occurring. Option A, node/cpu/usage_time, is not correct because it reflects aggregate CPU usage across the entire node, not per-container throttling behavior, so it cannot isolate a specific GKE workload's quota-induced throttling. Option C, container/memory/usage_bytes, is not correct because it measures memory consumption, which is unrelated to CPU throttling. Option E, container/accelerator/duty_cycle, is not correct because it pertains to GPU/accelerator utilization, not CPU quota throttling.

Answer

node/cpu/usage_time

Answer

container/memory/usage_bytes

Answer

container/accelerator/duty_cycle

Question 9

Which THREE approaches can help reduce egress costs while improving performance for a multi-region application using Cloud Load Balancing? (Choose 3)

Accepted Answer

Optimize data transfer by compressing responses.. Option A is correct because compressing responses (e.g., gzip or Brotli) shrinks the payload size that must leave the origin and traverse the network, directly lowering egress bytes billed and reducing transfer time for users. Option B is correct because internal load balancers keep inter-region traffic on Google's private network and, when combined with internal passthrough or internal application load balancing, avoid unnecessary public egress charges while routing traffic efficiently between regions. Option D is correct because Cloud CDN caches content at edge locations close to users, so repeated requests are served from the edge cache instead of the origin, cutting origin egress costs and improving latency. Option C is not correct because adding instances increases capacity and availability but does not by itself reduce egress charges or improve network performance. Option E is not correct because Premium Tier networking improves routing and latency but typically costs more than Standard Tier and does not directly reduce egress costs.

Answer

Increase the number of instances in each region.

Answer

Use premium tier networking for lower latency.

Question 10

Your company runs a multi-region e-commerce platform on Google Kubernetes Engine (GKE) with services in us-central1 and europe-west1. The application uses a global external HTTP(S) load balancer with Cloud CDN for static assets. Recently, users in Asia report that product images take 5-10 seconds to load, while users in the US and Europe experience sub-second load times. You check the Cloud CDN cache hit ratio and see it is 95% globally. You also notice that the images are served from a backend bucket in us-central1. The load balancer uses the default routing configuration. Your team has implemented client-side caching with Cache-Control headers set to public, max-age=3600. What is the most likely cause of the high latency for Asian users?

Accepted Answer

The cache hit ratio is too low for Asian users due to different traffic patterns.. While Cloud CDN has edge caches in Asia (e.g., Tokyo, Mumbai, Singapore), the global cache hit ratio of 95% does not guarantee a high hit ratio for Asian users. If their traffic patterns differ (e.g., first requests for new products), they may experience cache misses, forcing requests to the backend bucket in us-central1, resulting in high latency. Options A and B are not likely: the load balancer uses premium tier by default, and max-age=3600 is reasonable. Option C is incorrect because Cloud CDN does have edge caches in Asia.

Answer

The load balancer is not using premium tier networking, so traffic from Asia takes a longer path.

Answer

Client-side caching with max-age=3600 is too short, causing frequent revalidation.

Answer

Cloud CDN does not have edge caches in Asia, so requests are served from the nearest available edge location, which may be far from users.

Question 11

A company runs a critical application on Compute Engine instances behind a TCP/UDP Network Load Balancer. They notice intermittent high latency for a subset of users. The application logs show no errors, and instance CPU is below 50%. Which next step is most effective to diagnose the latency?

Accepted Answer

Analyze Cloud Monitoring metrics for the load balancer, including backend latency and request counts.. Cloud Monitoring provides detailed metrics for TCP/UDP Network Load Balancers, including backend latency and request counts, which directly help identify whether the latency originates from the load balancer itself or the backend instances. Since instance CPU is below 50% and application logs show no errors, the issue is likely at the network or load balancer level, and these metrics offer the most targeted diagnostic data without changing the architecture.

Answer

Increase the number of instances behind the load balancer.

Answer

Enable VPC Flow Logs and analyze for dropped packets.

Answer

Switch to an HTTP(S) Load Balancer for better visibility.

Question 12

A team uses Spanner for a global database. They notice increased read latency and high CPU utilization on some nodes. The workload is read-heavy with occasional writes. Which action is most likely to improve performance?

Accepted Answer

Add more nodes to the instance and ensure read requests are distributed evenly.. In a read-heavy Spanner workload with high CPU utilization on some nodes, adding more nodes and ensuring read requests are distributed evenly (Option D) directly addresses the bottleneck by increasing the instance's compute capacity and spreading the load across all nodes. Spanner's architecture uses a shared-nothing design where each node handles a portion of the data and traffic; uneven distribution can cause hot spots. Adding nodes scales out processing power, and ensuring even distribution (e.g., via proper key design or using Spanner's built-in load balancing) reduces latency and CPU spikes on individual nodes.

Answer

Create read-only replicas in each region.

Answer

Split the most frequently read tables into smaller tables.

Answer

Increase the number of nodes in the instance.

Question 13

A company runs a stateful workload on Compute Engine with local SSDs. They need to improve disk I/O performance without changing the instance type. Which THREE actions should they take?

Accepted Answer

Stripe data across multiple local SSD volumes using RAID 0.. Option B is correct because striping data across multiple local SSD volumes with RAID 0 aggregates the throughput and IOPS of each device, directly increasing disk I/O performance without changing the instance type. Option C is correct because mounting an SSD-appropriate filesystem such as ext4 with noatime and nodiratime reduces unnecessary metadata write operations, lowering I/O overhead and improving performance. Option D is correct because placing the instance in the same zone as the application that accesses the disks minimizes network latency and avoids cross-zone traffic, which is important for I/O-bound stateful workloads. Option A is not correct because migrating to persistent SSD changes the storage type and does not leverage the higher performance of local SSDs, and durability is not the stated goal. Option E is not correct because local SSD encryption is handled by Google's infrastructure and enabling encryption does not reduce I/O overhead or improve performance.

Answer

Migrate to persistent SSD for better durability.

Answer

Enable encryption for the local SSDs to reduce I/O overhead.

Question 14

Refer to the exhibit. An App Engine application returns 504 errors. The application calls an external API and processes the result. Which change is most likely to resolve the errors?

Accepted Answer

Increase the App Engine request timeout in app.yaml to 120 seconds.. A 504 error from App Engine indicates the request exceeded the timeout limit before the application could respond. The default App Engine request timeout is 60 seconds, and since the application calls an external API and processes the result, the total time may exceed this limit. Increasing the request timeout in app.yaml to 120 seconds allows the application more time to complete the external API call and processing, resolving the 504 error.

Answer

Reduce the idle timeout in the scaling settings.

Answer

Change the scaling type from automatic to manual.

Answer

Increase the number of instances to handle the load.

Question 15

Your team deploys a microservice on Google Kubernetes Engine (GKE) that serves an API with low latency requirements. Users report that the API occasionally times out during peak hours. You check the GKE metrics and see that CPU utilization is below 50% but memory is near 100% on the nodes. What is the most likely cause and what should you do?

Accepted Answer

The application is memory-constrained; increase memory resource limits for the pod.. The symptoms—low CPU utilization but near 100% memory usage on nodes, with API timeouts during peak hours—indicate that the application is hitting memory limits. When a pod exceeds its memory request, the kernel can OOM-kill it, causing request failures and timeouts. Increasing the memory resource limits for the pod allows it to allocate more heap or cache, preventing OOM kills and stabilizing latency.

Answer

The nodes are under-provisioned; add more nodes to the cluster.

Answer

The application is CPU-bound; increase CPU resource limits for the pod.

Answer

The network bandwidth is insufficient; increase the machine type for nodes.

Question 16

You created the above alert policy to detect high CPU utilization in your GKE cluster. However, you are receiving too many false positive alerts. What is the most likely reason?

Accepted Answer

The crossSeriesReducer is set to REDUCE_SUM, which sums CPU across containers, so a namespace with many containers can trigger the alert even if each container uses less than 80%.. The crossSeriesReducer set to REDUCE_SUM aggregates CPU utilization across all containers in a namespace. This means that even if each container uses only 20% CPU, a namespace with five containers would show a total of 100%, triggering the alert when the threshold is 0.8 (80%). This causes false positives because the alert fires on the sum, not on individual container utilization.

Answer

The threshold value of 0.8 is too low; it should be 0.9 for production.

Answer

The duration of 300 seconds (5 minutes) is too short; it should be longer to avoid transient spikes.

Answer

The filter does not specify a specific namespace, causing alerts from all namespaces.

Question 17

Your company runs a web application on Compute Engine behind a global HTTP(S) Load Balancer. You want to improve performance for users in Europe. You have already enabled Cloud CDN. What is the next best action to reduce latency?

Accepted Answer

Add Compute Engine instances in a European region (e.g., europe-west1) and add them to the load balancer's backend.. Adding Compute Engine instances in a European region (e.g., europe-west1) and adding them to the load balancer's backend reduces latency by bringing the origin server physically closer to European users. Even with Cloud CDN enabled, cache misses still require a round trip to the backend; having a backend in Europe minimizes that distance. The global HTTP(S) Load Balancer automatically routes requests to the closest healthy backend, so this directly improves performance for users in Europe.

Answer

Switch to a regional load balancer to route traffic more efficiently.

Answer

Configure a multi-regional Cloud CDN and enable cache warming for popular content.

Answer

Increase the machine type of existing instances to improve processing speed.

Question 18

You are troubleshooting a performance issue with a Compute Engine instance that is part of a managed instance group serving a web application. Users report intermittent high latency. You run the command shown in the exhibit. Based on the output, what is the most likely cause of the performance issue?

Accepted Answer

The instance is under-provisioned for CPU.. The output shows high CPU utilization (e.g., 95%+), which directly correlates with the reported intermittent high latency. In a managed instance group, if the instance is under-provisioned for CPU, it cannot handle the workload spikes, causing queuing and increased response times. This is the most common cause of performance degradation in Compute Engine instances serving web applications.

Answer

The instance is hitting the network egress bandwidth limit.

Answer

The service account lacks the necessary scopes for Cloud Monitoring and Cloud Trace.

Answer

The boot disk is too small, causing I/O contention.

Question 19

You are a DevOps engineer at a media streaming company. Your application runs on Google Kubernetes Engine (GKE) and serves video content to users worldwide. The application uses a microservices architecture with a frontend service that handles user requests and a backend transcoding service that converts video files. Recently, you noticed that the transcoding service is causing performance bottlenecks during peak hours, leading to increased latency for users. You have enabled Cloud Monitoring and Cloud Trace and observed that the transcoding service's CPU utilization is consistently above 90% during peak times, and the queue of video transcoding tasks is growing. The current deployment has 5 replicas of the transcoding service with no autoscaling. You need to optimize the performance of the transcoding service to reduce latency. Your company has a limited budget and wants to minimize costs. What should you do?

Accepted Answer

Refactor the frontend to push transcoding tasks to a Cloud Pub/Sub topic, and create a separate deployment of workers that subscribe to the topic and perform transcoding. Configure HPA on the worker deployment based on the Pub/Sub subscription backlog.. It decouples the transcoding workload from user-facing requests using Cloud Pub/Sub, allowing the worker deployment to scale independently based on the backlog of tasks. This pattern reduces latency by preventing the frontend from being blocked by transcoding, and HPA on Pub/Sub backlog ensures cost-efficient scaling only when demand increases, aligning with the limited budget.

Answer

Enable Horizontal Pod Autoscaling (HPA) on the transcoding service based on CPU utilization, targeting 70% utilization.

Answer

Upgrade the transcoding service to a larger machine type with more CPU and memory.

Answer

Increase the number of replicas of the transcoding service to 10 and keep it static.

Question 20

Match each monitoring concept to its purpose.

Accepted Answer

Latency: Time taken to serve a request. The four golden signals are Latency, Traffic, Errors, and Saturation. Correct matches are A, B, C, D. Common confusions swap Error and Traffic definitions.

Answer

Latency: Rate of failed requests

Answer

Errors: Demand placed on the system