PMLE Serving and Scaling Models Practice Test 4 — 15 Questions

Question 1

You are deploying a scikit-learn model to a Vertex AI endpoint using a custom container. The model artifacts are stored in a Cloud Storage bucket. During testing, predictions succeed, but you notice that each request takes several seconds because the model is loaded from Cloud Storage on every request. You want to minimize latency without changing the model. What should you do?

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

Accepted Answer

Configure the custom container to download the model artifacts to local storage during container startup, and load the model into memory before serving requests.. The latency issue stems from loading the model from Cloud Storage on every request. The correct approach is to download the model artifacts once during container startup and keep the model in memory for subsequent predictions. This eliminates repeated network calls and leverages the container's lifecycle to prepare the model before serving traffic, which is the recommended pattern for custom containers on Vertex AI.

Answer

Increase the endpoint's minimum replica count to ensure that at least one replica is always warm and has the model cached.

Answer

Add a Cloud Storage FUSE mount to the container so the model files are available as a local filesystem, and load the model from the mount on each request.

Answer

Enable request-response logging on the endpoint to identify slow requests, then optimize the model's inference code to reduce computation time.

Question 2

You are deploying a PyTorch image classification model to a Vertex AI endpoint. The model's container reads images from Cloud Storage at prediction time, and each prediction takes about 800 ms. You want to minimize cold-start latency when the endpoint scales out. Which deployment setting should you configure?

Accepted Answer

Deploy the model with minReplicaCount equal to the expected peak concurrency and keep those replicas warm.. Cold-start latency on Vertex AI endpoints comes from container initialization and model loading, not from the inference compute itself. To avoid it, capacity must be provisioned in advance. Setting minReplicaCount to the expected peak concurrency keeps enough replicas warm so that incoming requests are served immediately, without waiting for new replicas to start and load the model. Other settings influence isolation, health checks, or per-request speed but do not pre-warm replicas.

Answer

Configure a container health route and set the startup probe to fail fast.

Answer

Set minReplicaCount to 1 and maxReplicaCount to 3 on the DeployedModel.

Answer

Enable dedicated endpoint and configure a higher machine type with more vCPUs.

Question 3

You are serving a scikit-learn model on a Vertex AI Endpoint. Predictions must be logged for auditing, but the input features contain personally identifiable information (PII). Your security team requires that PII be removed before logs are written to Cloud Logging. You want to keep using the default prebuilt scikit-learn container. What should you do?

Accepted Answer

Deploy a custom prediction routine that redacts PII from the input before calling the model, and enable request-response logging.. To keep logs but exclude PII, sanitization must occur before the payload is written to Cloud Logging. A custom prediction routine provides a hook to redact or mask PII fields in the incoming request, so the logged request contains no sensitive data. Filters and partial logging do not satisfy the requirement to remove PII before logging.

Answer

Use Vertex AI Feature Store to store features and disable logging of raw requests.

Answer

Enable request-response logging on the endpoint and configure a Cloud Logging exclusion filter to drop PII fields.

Answer

Configure the endpoint to log only response bodies and not request bodies.

Question 4

You are serving a TensorFlow model on a Vertex AI endpoint with autoscaling enabled. The endpoint is configured with minReplicaCount=1 and maxReplicaCount=10. You observe that during sudden traffic spikes, the endpoint returns 429 (Too Many Requests) errors for several minutes before scaling out. You want to reduce these errors while keeping costs low during idle periods. What should you do?

Accepted Answer

Configure the endpoint to use a higher minReplicaCount during peak hours via scheduled scaling, and keep a low minReplicaCount during off-peak hours.. To reduce 429 errors during sudden spikes while keeping costs low during idle, use scheduled scaling to increase minReplicaCount during predictable peak hours. This pre-warms replicas, so the endpoint can handle traffic immediately without waiting for autoscaling to react. Reactive autoscaling alone may be too slow for sudden spikes, but scheduled scaling provides capacity ahead of time.

Answer

Configure the endpoint to use a larger machine type for the replicas to handle more requests per replica.

Answer

Set a higher minReplicaCount and enable autoscaling with a lower target utilization to trigger scale-out earlier.

Answer

Increase the minReplicaCount to 3 and set a higher target utilization for autoscaling.

Question 5

A financial institution is deploying a fraud detection model to a Vertex AI endpoint. The model is a custom PyTorch model that requires GPU acceleration. The endpoint must handle bursts of traffic while minimizing cost. They have configured minReplicaCount=1, maxReplicaCount=10, and machine type n1-standard-4 with one NVIDIA T4 GPU. Which TWO additional configurations should they apply to optimize for both performance and cost? (Choose two.)

Accepted Answer

Enable autoscaling based on GPU duty cycle with a target of 70%.. Autoscaling based on GPU duty cycle ensures that replicas are added when the GPU is busy, which is critical for GPU-accelerated models. Logging to BigQuery provides insights to fine-tune autoscaling settings over time. Together, they optimize performance during bursts and reduce cost by avoiding unnecessary replicas. The other options are either prerequisites, irrelevant, or detrimental to cost efficiency.

Answer

Configure the model server to use a batch size of 1 to minimize latency.

Answer

Set the endpoint to use a dedicated public IP address for secure access.

Answer

Use a custom container that includes the NVIDIA CUDA libraries and set the environment variable to enable GPU support.

Question 6

A media company runs a Vertex AI endpoint that returns personalized article rankings. Peak-hour latency has crept above the service-level objective, and profiling shows the model's forward pass dominates response time. The model uses a transformer with a fixed maximum sequence length far larger than most real requests. Which two actions should you take to reduce serving latency while preserving ranking quality? (Choose two.)

Accepted Answer

Set the model's maximum sequence length to match the observed request distribution and truncate the rare longer inputs at inference time.. The profile points at the forward pass, so the effective levers are making that computation cheaper. Matching the maximum sequence length to the real input distribution removes wasted attention work, and compiling the graph into an optimized runtime reduces per-operation overhead. Both changes leave the model's learned behavior intact for typical inputs, so ranking quality is preserved while latency drops measurably.

Answer

Lower the endpoint's minReplicaCount so fewer replicas serve traffic and each replica's CPU cache stays warmer between requests.

Answer

Enable request-response logging on the endpoint and add a Cloud Monitoring dashboard for latency percentiles so the team can watch the metric during peak hours.

Answer

Configure the endpoint's deployed model to enable dynamic batching so concurrent requests are grouped into a single forward pass.

Question 7

A retail company is using Vertex AI Endpoints to serve a demand forecasting model. They notice that during peak shopping hours, prediction latency increases significantly. They want to reduce latency without increasing costs unnecessarily. Which TWO actions should they take? (Choose two.)

Accepted Answer

Enable autoscaling with a higher maximum replica count and set a lower target CPU utilization threshold.. Enabling autoscaling with a higher max replica count and lower CPU threshold allows dynamic scaling to meet demand, while a minimum replica count ensures warm capacity. Together, they reduce latency during peaks without excessive cost during lulls.

Answer

Increase the machine type of the endpoint's replicas to a larger instance with more vCPUs and memory.

Answer

Configure the endpoint to use a smaller batch size for predictions.

Answer

Reduce the model's input feature vector size by removing less important features.

Question 8

You are optimizing the cost of a Vertex AI endpoint that serves a model with variable traffic. The endpoint currently uses a single machine type and autoscaling. You want to reduce cost without degrading latency during peak hours. (Choose two.)

Accepted Answer

Configure autoscaling based on a custom metric that reflects request queue depth.. Scale-to-zero eliminates idle replica costs during periods of no traffic, and autoscaling on a custom metric like queue depth ensures replicas are added proactively to maintain latency during spikes. Together, these reduce cost by matching capacity to demand while preserving performance during peaks.

Answer

Use a smaller machine type for the endpoint.

Answer

Increase the maxReplicaCount to handle peak traffic.

Answer

Deploy the model to a Dedicated Endpoint with a fixed number of replicas.

Question 9

A team is serving a large language model on a Vertex AI endpoint with a single GPU per replica. They observe that GPU memory is exhausted under load and that throughput drops sharply when many requests arrive together. They want to improve GPU utilization and throughput without changing the model. Which TWO changes should they make? (Choose two.)

Accepted Answer

Increase the number of GPUs per replica and enable model parallelism so the model is sharded across GPUs.. GPU memory exhaustion within a replica is addressed by spreading the model across more GPUs, while the throughput drop under concurrent load is addressed by batching requests into single forward passes. Together, model parallelism and request batching raise utilization and throughput without modifying the model itself.

Answer

Enable request-response logging to capture the payloads that cause memory exhaustion.

Answer

Switch the endpoint to batch prediction so requests are processed offline in large jobs.

Answer

Reduce the endpoint's maxReplicaCount to lower contention for GPU memory.

Question 10

A team is using Vertex AI Endpoints to serve a model that performs image classification. They notice that during peak hours, the endpoint's P99 latency increases significantly, and some requests time out. The model is deployed on a GPU machine type. They want to reduce latency and improve reliability without changing the model. Which TWO actions should they take? (Choose two.)

Accepted Answer

Increase the number of replicas and configure autoscaling based on GPU utilization.. To reduce latency and improve reliability under peak load without changing the model, the team should enable request batching to improve GPU utilization and add replicas with GPU-based autoscaling to distribute the load. Batching processes multiple requests together, increasing throughput, while autoscaling ensures enough capacity to handle traffic spikes. These actions are configuration changes that directly address the bottleneck.

Answer

Enable Vertex AI Model Monitoring to detect latency anomalies and automatically scale the endpoint.

Answer

Reduce the model's input image resolution to decrease inference time.

Answer

Switch to a larger GPU machine type with more memory and compute cores.

Question 11

You are deploying a model to a Vertex AI endpoint and need to configure private networking so that prediction requests do not traverse the public internet. You also want to control which VPC networks can reach the endpoint. Which two configurations should you implement? (Choose two.)

Accepted Answer

Enable Private Service Connect on the endpoint and create a forwarding rule in the consumer VPC.. Private access to a Vertex AI endpoint is achieved with Private Service Connect. The producer, Google, publishes the service, and the consumer creates a Private Service Connect endpoint in their VPC with an internal IP address. A forwarding rule and endpoint in the consumer VPC route traffic privately. This keeps prediction requests off the public internet and lets the consumer control access through VPC firewall rules and routing, satisfying both private connectivity and network-level access control.

Answer

Configure VPC Service Controls with an access policy that includes the project and restricts the endpoint to authorized networks.

Answer

Set the endpoint to use a private IP address and configure Cloud NAT for egress.

Answer

Enable Cloud VPN between the consumer VPC and the Vertex AI service project.

Question 12

A global e-commerce company uses a Vertex AI Endpoint to serve a product ranking model. The model is deployed with minReplicaCount=2 and maxReplicaCount=20. During a flash sale, traffic increases tenfold within seconds. The endpoint's autoscaler takes several minutes to add replicas, causing timeouts. The team wants to reduce the time to scale out. Which action should they take?

Accepted Answer

Configure the endpoint to use a custom autoscaling metric that reflects request concurrency, and set a lower threshold.. The autoscaler's reaction time depends on the metric it monitors. CPU utilization can lag during sudden spikes, so using a custom metric like request concurrency with a lower threshold can trigger scaling sooner. This reduces the time to add replicas and prevents timeouts during flash sales. Other options either do not address the detection delay or add unnecessary complexity.

Answer

Deploy the model to multiple endpoints and use a load balancer to distribute traffic.

Answer

Increase the endpoint's minReplicaCount to 20 to match the peak traffic.

Answer

Reduce the model's container image size to speed up replica startup.

Question 13

You are deploying a large language model (LLM) to a Vertex AI endpoint for real-time text generation. The model is hosted in a custom container and requires a GPU. You need to minimize latency for each request while keeping costs under control. The model uses a batch size of 1 for inference. What should you do?

Accepted Answer

Use a GPU with more memory to allow the model to be loaded entirely into GPU memory, reducing CPU-GPU transfers.. For real-time LLM inference, the primary latency bottleneck is often the transfer of model parameters between CPU and GPU memory. Ensuring the model fits entirely in GPU memory eliminates these transfers, allowing each inference to run without memory swapping. This requires selecting a GPU with enough memory to hold the model. While dynamic batching improves throughput, it adds latency, so it's not suitable for latency-sensitive applications.

Answer

Enable dynamic batching on the endpoint to combine multiple requests into a single batch for GPU processing.

Answer

Set the endpoint to use a CPU-only machine type and rely on optimized CPU inference libraries.

Answer

Configure the endpoint to use multiple replicas with smaller GPUs to distribute the load across more devices.

Question 14

A healthcare company has a Vertex AI endpoint serving a TensorFlow model for medical image classification. The model is deployed on a n1-standard-4 machine with no GPU. During peak hours, the endpoint's latency increases significantly, and the autoscaler adds replicas but the latency remains high. Monitoring shows that CPU utilization is around 40% even at peak. You need to reduce latency without increasing the number of replicas. What should you do?

Accepted Answer

Enable GPU acceleration by redeploying the model on a machine type with an NVIDIA T4 GPU.. The low CPU utilization suggests the bottleneck is not CPU but likely the model's computational demands, which can be accelerated by a GPU. Redeploying the TensorFlow model on a GPU-enabled machine will reduce per-request latency without increasing replica count. Increasing machine size or replicas does not address the root cause, and quantization may require model changes and accuracy trade-offs.

Answer

Change the machine type to one with more vCPUs and memory, and redeploy the model.

Answer

Optimize the model by quantizing it to reduce its size and computational requirements.

Answer

Increase the number of replicas beyond the current maximum to distribute the load.

Question 15

A company has a Vertex AI endpoint serving a model that performs image classification. They want to minimize cost during nights and weekends when traffic is near zero, but they still need the endpoint to respond to occasional requests within a few seconds. What should they do?

Accepted Answer

Set minReplicaCount=0 and maxReplicaCount to a small value so the endpoint scales to zero when idle and scales up on demand.. A near-zero-traffic pattern with tolerance for a few seconds of latency is the ideal case for scale-to-zero. Setting the minimum replica count to zero eliminates idle cost, and the endpoint's autoscaler brings a replica up when a request arrives, satisfying the occasional-response requirement.

Answer

Keep minReplicaCount=2 and rely on a committed use discount to offset the idle cost.

Answer

Delete the endpoint at night and recreate it each morning before traffic resumes.

Answer

Move the model to a Vertex AI batch prediction job that runs every hour and caches the results.