PCDOE Optimizing service performance Practice Test 4 — 15 Questions

Question 1

A company runs a web application on Compute Engine behind a regional HTTP Load Balancer. Users report slow page load times during peak hours. CPU utilization on instances is under 60%, but network egress is near the instance's bandwidth limit. Which action should the engineer take?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Use a larger machine type with higher network throughput.. The bottleneck is network egress bandwidth, not CPU. Increasing the instance size to a machine type with higher network throughput (e.g., n2-highmem-4 vs. n2-standard-2) directly raises the per-instance egress cap, alleviating the bandwidth limit. Option D is correct because it addresses the root cause—insufficient network I/O capacity per instance.

Answer

Increase the number of instances in the instance group.

Answer

Switch to a global external HTTP Load Balancer.

Answer

Enable Cloud CDN.

Question 2

A DevOps team uses Cloud Run for a containerized application that processes real-time financial data. The service has a concurrency setting of 80, and instances are scaled based on CPU usage. During market volatility, the service experiences high latency and some requests timeout. Cloud Monitoring shows that the average CPU utilization is 40%, but the instance count spikes to the maximum allowed. What is the most likely cause?

Accepted Answer

The concurrency setting is too low, causing many instances to be created.. With a concurrency setting of 80, each instance can handle up to 80 simultaneous requests. However, if the actual request rate exceeds 80 per instance, Cloud Run will spin up new instances. During market volatility, the request volume spikes, causing the instance count to hit the maximum even though average CPU utilization is only 40%. This indicates that the concurrency limit is too low for the burst traffic, forcing excessive instance creation and leading to high latency and timeouts due to instance startup overhead.

Answer

The max instances limit is set too low, causing requests to queue.

Answer

The service uses too much memory, causing cold starts.

Answer

The CPU utilization target for autoscaling is set too high, causing slow scaling.

Question 3

A company has a stateful application deployed on a GKE cluster with stateful sets using persistent volumes. The application is experiencing higher than expected latency for write operations. The team uses SSDs for persistent disks. Cloud Monitoring shows high disk queue depth on the nodes where the stateful pods are scheduled. Which of the following is the most effective optimization?

Accepted Answer

Enable disk caching on the persistent disks.. Enabling read/write caching on persistent disks reduces write latency by buffering writes to the local instance's SSD before acknowledging them to the application. This directly addresses the high disk queue depth observed in Cloud Monitoring, as caching absorbs bursty write I/O and lowers queue depth. For stateful workloads on GKE with SSDs, disk caching is a standard optimization to improve write performance without changing the underlying disk type.

Answer

Configure a separate node pool with local SSDs for the stateful workloads.

Answer

Increase the number of replicas of the stateful set.

Answer

Use regional persistent disks for higher throughput.

Question 4

A media streaming service uses Cloud Storage to store video files and serves them via Cloud CDN. Users in Asia report buffering issues. The team notices that the cache hit ratio is low in that region. The origin is a single Cloud Storage bucket in us-central1. Which set of actions would best improve performance for Asian users?

Accepted Answer

Create a new Cloud Storage bucket in an Asian region and use dual-region bucket with Cloud CDN.. Creating a new Cloud Storage bucket in an Asian region and using a dual-region bucket with Cloud CDN reduces latency by serving content from a geographically closer origin, improving cache hit ratios for Asian users. Cloud CDN caches content at edge locations, but if the origin is far away (us-central1), the first miss still incurs high latency. A dual-region bucket provides a local origin for cache misses, significantly reducing round-trip time.

Answer

Use Cloud Load Balancing with Cloud Armor to protect the origin.

Answer

Enable HTTP/2 on Cloud CDN and increase the TTL for video content.

Answer

Configure a custom domain on Cloud CDN with SSL and enable request collapsing.

Question 5

A company deploys a microservices application on Google Kubernetes Engine (GKE). They notice increased latency during peak hours. The application uses a Cloud SQL database for state. The team wants to optimize service performance. What should they do first?

Accepted Answer

Enable Cloud SQL connection pooling.. The primary bottleneck during peak hours for a microservices application using Cloud SQL is often the number of database connections. Connection pooling reuses a fixed set of connections, reducing the overhead of establishing new connections and preventing connection exhaustion, which directly addresses increased latency without requiring architectural changes.

Answer

Implement a caching layer with Memorystore.

Answer

Move the database to Cloud Spanner.

Answer

Increase the number of replicas in the GKE deployment.

Question 6

A company serves static content using a global HTTP(S) load balancer with Cloud CDN. They want to maximize the cache hit ratio. Which two actions should they take?

Accepted Answer

Set Cache-Control: public, max-age=31536000.. Option B is correct because setting Cache-Control: public, max-age=31536000 explicitly marks the static content as cacheable by Cloud CDN and shared caches for up to one year, which maximizes cache hit ratio by allowing responses to be reused across many users. Option C is correct because configuring the Cloud CDN cache key to include only the URL and host (rather than headers, cookies, or query strings) ensures that requests for the same resource map to the same cached object, increasing the likelihood of cache hits. Option A is not appropriate because signed URLs are used to restrict access to content and typically add unique query parameters or signatures that fragment the cache key, reducing hit ratio. Option D is incorrect because Cache-Control: private prevents shared caches like Cloud CDN from storing the response, which would eliminate caching benefits. Option E is incorrect because enabling Identity-Aware Proxy requires authentication and typically bypasses or reduces CDN caching for protected resources, lowering the cache hit ratio.

Answer

Use signed URLs for all requests.

Answer

Set Cache-Control: private.

Answer

Enable Identity-Aware Proxy (IAP) for the backend.

Question 7

Refer to the exhibit. A team uses these Compute Engine instances to run a batch processing job. The job frequently gets killed on instance-3. What is the most likely cause?

Accepted Answer

Instance-3 is preemptible and gets terminated by Google Cloud.. Instance-3 is a preemptible VM, which means Google Cloud can terminate it at any time if its resources are needed elsewhere. Preemptible instances have a maximum runtime of 24 hours and are subject to termination with a 30-second warning. The batch processing job being frequently killed on instance-3 is a classic symptom of preemptible VM termination, not a resource exhaustion or disk issue.

Answer

Instance-3 has insufficient CPU.

Answer

Instance-3 has a corrupted disk.

Answer

Instance-3 has insufficient memory.

Question 8

A company runs a production web application on Google Compute Engine behind an HTTP(S) load balancer. The application is deployed across multiple managed instance groups in three regions (us-east1, europe-west1, asia-east1). Recently, users report slow page load times. Monitoring shows that CPU utilization on instances is consistently low (around 30%) but memory usage is high (over 80%). The application uses a self-managed in-memory cache per instance to store session data and frequently accessed objects. The team is considering adding more instances to the instance groups to distribute the load. However, they notice that the load balancer's latency is spiking and the cache hit ratio is low. What is the most likely issue and what should the engineer do?

Accepted Answer

Migrate to a managed in-memory cache like Memorystore for Redis to serve as a centralized cache shared by all instances.. The low cache hit ratio and high memory usage indicate that each instance's self-managed in-memory cache is fragmented and inefficient, as session data and frequently accessed objects are not shared across instances. This forces the load balancer to repeatedly fetch data from the backend, causing latency spikes. Migrating to a centralized managed cache like Memorystore for Redis eliminates per-instance cache duplication, improves cache hit ratio, and reduces load balancer latency by serving data from a single, consistent cache.

Answer

Add more instances to the instance groups to increase total memory capacity.

Answer

Increase the machine type of instances to have more memory per instance (e.g., n1-highmem-4).

Answer

Enable autoscaling based on memory utilization.

Question 9

A media company runs a GKE cluster hosting a video transcoding service. The transcoding pods are CPU-intensive and have been experiencing frequent OOMKilled events and CPU throttling. You need to optimize pod performance and ensure stable operation. What should you do?

Accepted Answer

Increase the CPU and memory limits for the transcoding pods and set resource requests equal to limits.. The transcoding pods are experiencing OOMKilled events and CPU throttling because their resource limits are too low. Setting requests equal to limits guarantees the pods receive the requested resources and prevents CPU throttling, while higher limits avoid OOMKilled events. This ensures stable performance for CPU-intensive workloads.

Answer

Configure a Horizontal Pod Autoscaler (HPA) to scale the transcoding pods based on CPU utilization.

Answer

Enable Vertical Pod Autoscaler (VPA) in recommendation mode to adjust resource requests automatically.

Answer

Use node auto-provisioning to add nodes with more CPU and memory to the cluster.

Question 10

Your team runs a stateful application on GKE using PersistentVolumes backed by Compute Engine persistent disks. During peak load, the application experiences high I/O latency. You want to optimize storage performance. Which two actions should you take? (Choose two.)

Accepted Answer

Increase the size of the persistent disk to gain additional baseline IOPS and throughput.. Switching to SSD persistent disks and increasing disk size are both effective ways to improve I/O performance for persistent disks on GKE. SSD disks provide higher IOPS and lower latency, while larger disks scale performance limits. Read-only access, disruption budgets, and Local SSDs do not address the need for durable, high-performance storage for a stateful application.

Answer

Enable read-only many access mode on the PersistentVolume to allow multiple pods to read concurrently.

Answer

Use a Local SSD instead of a persistent disk for the application's data.

Answer

Configure a PodDisruptionBudget to prevent pods from being evicted during high I/O.

Question 11

A retail company runs a Java service on GKE Standard. During flash sales, pods are terminated by the kubelet with exit code 137 even though node CPU and memory show headroom, and the service's p99 latency spikes. The containers are configured with requests of 500m CPU / 512Mi memory and no limits. You need to stop the terminations while preserving burst behavior and keeping the pods schedulable on the existing node pool. What should you do?

Accepted Answer

Increase the memory request to a realistic steady-state value and set a memory limit above it to allow bursting.. Exit code 137 signals SIGKILL, which in GKE almost always means the container exceeded a memory boundary, either a configured limit or the node's allocatable capacity. The fix is to size the memory request to genuine steady-state consumption so the scheduler places pods correctly, then set a limit above that value so short bursts are absorbed instead of triggering the OOM killer. CPU limits should be avoided because throttling degrades latency.

Answer

Configure a PodDisruptionBudget with minAvailable set to the replica count so pods cannot be terminated.

Answer

Add a container memory limit of 512Mi and keep the existing memory request unchanged.

Answer

Add a CPU limit of 500m and a memory limit of 512Mi so the pod is fully constrained and predictable.

Question 12

A retail company runs its order-processing service on GKE. During flash sales, the service's latency spikes and the operations team suspects a suboptimal CPU allocation. They want to ensure the container is scheduled with a guaranteed CPU allocation and is treated as high-priority by the scheduler. Which action should the team take?

Accepted Answer

Set the CPU request equal to the CPU limit and assign a high-priority class.. To guarantee CPU allocation and ensure high scheduling priority, the Pod must be in the Guaranteed QoS class (CPU request equals limit) and have a high-priority class. This combination ensures the container gets dedicated CPU and is less likely to be evicted or preempted during resource contention. The other options either provide only partial benefits or are not valid GKE features.

Answer

Set the CPU limit equal to the CPU request in the Pod specification.

Answer

Enable CPU burst on the node pool.

Answer

Assign a higher Pod priority class to the Deployment.

Question 13

You are deploying a new version of a web application on Google Kubernetes Engine (GKE). You want to minimize downtime and ensure that only healthy pods receive traffic during the rollout. Which GKE feature should you use?

Accepted Answer

Rolling update strategy on the Deployment. A rolling update strategy on a Deployment allows you to update pods incrementally, ensuring that a minimum number of pods are available and that new pods are healthy before old ones are terminated. This, combined with readiness probes, minimizes downtime. HPA, PDB, and readiness probes alone do not orchestrate the rollout process in the same way.

Answer

Readiness probes

Answer

PodDisruptionBudget (PDB)

Answer

Horizontal Pod Autoscaler (HPA)

Question 14

A financial services firm runs a stateless API on a regional managed instance group behind an external Application Load Balancer. During month-end processing, the autoscaler adds instances, but p99 latency still climbs and the load balancer reports uneven distribution with several backends at high utilization. The instances use a 60-second connection draining timeout and the API holds long-lived HTTP/2 streams. Which change best improves backend balance and tail latency?

Accepted Answer

Switch the backend service balancing mode to RATE and set a max rate per instance so connections spread by request rate.. With HTTP/2, a small number of connections can carry many concurrent requests, so connection-based balancing leaves some backends idle while others are saturated. Switching the backend service to rate-based balancing lets the load balancer distribute by observed request rate, which rebalances new streams to less-loaded instances. This directly targets uneven utilization and tail latency without simply adding capacity.

Answer

Increase the connection draining timeout to 300 seconds so in-flight requests complete during scale-in events.

Answer

Enable Cloud CDN on the backend service so repeated API responses are served from edge caches.

Answer

Lower the autoscaler's target CPU utilization from 60 percent to 30 percent so more instances are added.

Question 15

A SaaS company runs a stateful analytics service on GKE Autopilot using PersistentVolumeClaims backed by the standard-rwo StorageClass. Users report that report generation slows dramatically when many tenants read the same dataset simultaneously, and Cloud Monitoring shows high disk read latency on the underlying Persistent Disk. You need to reduce read latency for this shared-read workload with minimal application change. (Choose two.)

Accepted Answer

Switch the PVCs to a StorageClass backed by Hyperdisk Throughput and increase the provisioned throughput.. The workload is a shared-read pattern where many tenants read the same dataset, so the fixes must either raise provisioned storage performance or let readers share one cached copy. Hyperdisk Throughput provides independently tunable throughput beyond what size-linked Persistent Disk offers, and a Filestore-backed read-only mount lets all reader pods share a single NFS dataset. Together they attack the bottleneck directly without application rewrites.

Answer

Set the pod's volumeMount to readOnly so the kernel caches pages and serves repeated reads from memory.

Answer

Increase the PVC size from 100Gi to 1Ti so the Persistent Disk automatically receives more IOPS.

Answer

Enable the GKE node auto-repair feature so degraded nodes are replaced before reads slow down.