PCDE Optimising Service Performance — 10 Questions

Question 1

Your GKE cluster is running a critical web application that experiences predictable traffic spikes during business hours. You want to minimize latency and avoid pod startup delays during scaling. The application uses CPU-intensive image processing. Which scaling strategy should you use?

Accepted Answer

Configure HPA with a minimum of 2 replicas and scale on CPU utilization; enable cluster autoscaler for node provisioning.. To avoid cold starts while ensuring pods can handle CPU spikes, you need a baseline of pods and dynamic scaling responsive to CPU. HPA with a minimum replicas of 2 ensures baseline capacity; HPA scales on CPU. Cluster autoscaler adds nodes if needed, but does not directly address pod startup delay. VPA adjusts resource requests, which can help but does not prevent cold starts. Using HPA alone with min replicas avoids pod creation latency.

Answer

Set a high number of static pods equal to peak traffic; use cluster autoscaler to add nodes.

Answer

Use VPA with updateMode: Auto to automatically adjust pod resources; enable cluster autoscaler to add nodes as required.

Answer

Deploy a CronJob to scale up replicas before business hours; rely on HPA to handle the rest.

Question 2

You are running a Cloud Run service that performs background image processing tasks after responding to HTTP requests. Users report that tasks are sometimes not completed. You have set CPU to always-on, but you notice the service's concurrency is set to the default of 80. What is the most likely cause of incomplete tasks?

Accepted Answer

The concurrency setting of 80 allows too many requests to share the instance's CPU, starving background tasks.. Cloud Run allows up to 80 concurrent requests per instance by default. If the concurrency is too high, the instance's CPU may be overwhelmed, causing background tasks to be preempted or not completed. Setting CPU to always-on ensures CPU is available but does not limit concurrency. Reducing concurrency to a lower value (e.g., 1) ensures each request gets dedicated CPU for background work.

Answer

The service should use Gen2 execution environment for better background processing.

Answer

The CPU always-on setting is not sufficient; you need to enable background CPU.

Answer

You need to increase the request timeout beyond 60 minutes.

Question 3

Your Cloud Spanner instance is experiencing high write latency and hot spots on a table that uses an auto-incrementing integer as the primary key. Which change would best mitigate the hot spots?

Accepted Answer

Change the primary key to a random UUID.. Monotonically increasing keys (like auto-incrementing integers) cause all writes to go to the same tablet (hot spot) in Cloud Spanner. Using a key with uniform distribution, such as a random UUID or a hash prefix, spreads writes across splits. Bit-reverse index is an alternative for sequential keys, but a random key is simpler. Interleaved tables and secondary indexes do not address the root cause.

Answer

Add a secondary index on the auto-incrementing column.

Answer

Increase the number of splits by raising the number of nodes.

Answer

Use interleaved tables to store related data together.

Question 4

Your organization uses Cloud SQL for PostgreSQL for a reporting application with read-heavy workloads. Queries are slow and you need to reduce load on the primary instance. You also need to ensure that all read queries from the reporting tool are isolated from the primary. What should you do?

Accepted Answer

Create a Cloud SQL read replica, and configure the reporting tool to connect to the replica's IP address.. Cloud SQL read replicas serve read traffic and offload the primary. For isolation, the reporting tool must connect to the replica's IP. Cloud SQL Auth Proxy is a secure tunnel, but does not provide connection pooling. PgBouncer can be used with proxies for connection pooling, but the question asks for isolation, which is achieved by pointing the reporting tool to the replica. The best answer is to create a read replica and configure the reporting tool to connect to it.

Answer

Create a Cloud SQL clone and point the reporting tool to the clone.

Answer

Enable connection pooling using Cloud SQL Auth Proxy and PgBouncer on the primary instance.

Answer

Configure the reporting tool to use the primary instance with a lower priority.

Question 5

You are deploying a batch job on Compute Engine that processes large datasets and can tolerate interruptions. The job runs for about 6 hours every night. Which option would minimize cost while ensuring the job completes within the 6-hour window?

Accepted Answer

Use preemptible VMs to get up to 80% discount, with checkpointing to resume if preempted.. Preemptible VMs are significantly cheaper (up to 80% discount) but can be terminated at any time. For fault-tolerant batch jobs, preemptible VMs are ideal; if preempted, the job can resume on another preemptible VM. Committed use discounts provide cost savings for steady-state usage (1- or 3-year commitment), but are not suitable for a nightly 6-hour job. Custom machine types or standard VMs would be more expensive.

Answer

Use standard VMs with committed use discounts for 1 year.

Answer

Use custom machine types to optimize cost.

Answer

Use sole-tenant nodes for dedicated hardware.

Question 6

Your Cloud Run service processes requests from an external API that sends a burst of up to 100 requests per second. You want to maximize throughput while minimizing instances. The service is CPU-bound. What configuration should you use?

Accepted Answer

Set concurrency to 1 to give each request full CPU, and enable CPU always-on.. For CPU-bound services, the default concurrency of 80 may be too high, causing resource contention and slowing down each request. Reducing concurrency to a lower value, such as 1, dedicates the entire CPU to each request, potentially improving throughput per instance. CPU always-on is needed for background tasks but not necessarily for request handling. Gen2 execution environment can help with higher memory but not CPU-bound throughput directly.

Answer

Use the Gen2 execution environment with higher memory.

Answer

Set concurrency to 1000 to handle burst efficiently.

Answer

Set concurrency to 80 (default) and CPU always-on.

Question 7

Your GKE cluster runs a stateful application that requires persistent storage. You want to use Vertical Pod Autoscaler (VPA) to optimize resource requests, but you notice that pods are being terminated and recreated when VPA updates resource recommendations. You want to avoid pod restarts. Which VPA updateMode should you use?

Accepted Answer

updateMode: Initial. VPA updateMode: Auto will evict and recreate pods to apply new resource limits. To avoid restarts, you can use Initial (applies only to new pods) or Off (no changes). The question asks to avoid pod restarts, so Initial is appropriate if you can accept that existing pods keep their old limits. Recreate is the same as Auto. Off disables VPA.

Answer

updateMode: Recreate

Answer

updateMode: Off

Answer

updateMode: Auto

Question 8

Your Pub/Sub subscription is not keeping up with the message publishing rate. The subscriber is a Cloud Run service that processes each message in about 2 seconds. You have already increased the number of subscribers to 10. What is the next best step to increase throughput?

Accepted Answer

Increase the acknowledgement deadline to 30 seconds.. If increasing subscribers does not help, the issue may be the acknowledgement deadline. If the deadline is too short, messages are redelivered before processing completes, causing duplicates and wasted work. Increasing the deadline (from default 10 seconds to something higher like 30 seconds) gives time to process. Flow control limits throughput. Ordering keys can reduce throughput.

Answer

Enable message ordering to ensure sequential processing.

Answer

Switch to a pull subscription instead of push.

Answer

Reduce the flow control max outstanding messages to 100.

Question 9

You want to ensure that a critical deployment on GKE has minimal downtime during rolling updates. You also want to ensure that at least 2 pods are always available. Which Kubernetes resource should you configure?

Accepted Answer

PodDisruptionBudget with minAvailable: 2. PodDisruptionBudget (PDB) specifies the minimum number or percentage of pods that must be available during voluntary disruptions like rolling updates. Setting minAvailable: 2 ensures at least 2 pods are running during updates. HPA, VPA, and cluster autoscaler do not control pod availability during updates.

Answer

Cluster autoscaler with minNodes: 2

Answer

HorizontalPodAutoscaler with minReplicas: 2

Answer

VerticalPodAutoscaler with updateMode: Auto

Question 10

Your Cloud Run service experiences cold starts that add 5 seconds of latency to user requests. You want to eliminate cold starts for a baseline traffic level of 10 requests per second. The service uses about 2 GiB of memory. What should you do?

Accepted Answer

Set minInstances to 2 to keep instances always warm.. Cloud Run cold starts can be eliminated by setting a minimum number of instances that are always warm. The minInstances flag ensures that the specified number of instances are always running and ready to serve requests. For a baseline of 10 RPS, you need to estimate how many instances are needed; but the question asks to eliminate cold starts, so setting minInstances to at least 1 (or more) will keep instances warm. Setting CPU always-on does not eliminate cold starts. Gen2 execution environment may reduce cold start time but not eliminate them.

Answer

Use the Gen2 execution environment to reduce cold start time.

Answer

Increase the concurrency to 1 to minimize instance count.

Answer

Set CPU always-on to keep the instance active.