PMLE question patterns

Climb the effort ladder: pre-trained API first, then BigQuery ML or AutoML, and custom training last.

Each rung down adds engineering effort that the requirement rules out.

Train with CREATE MODEL and score with ML.PREDICT inside BigQuery.

No export, no pipeline, and governance stays in the warehouse.

Call the Vision, Translation or Natural Language API.

Nothing to train when the labels are generic.

Train an AutoML model on a labelled managed dataset.

Pre-trained APIs can't learn custom classes, and AutoML avoids hand-building architectures.

Package the code and run a custom job with worker pools.

Only custom training exposes that level of control.

Use Vertex AI Search or RAG Engine over a data store.

Retrieval keeps knowledge current and cites sources. Tuning bakes in stale facts.

Supervised fine-tuning or LoRA on example prompt and response pairs.

Tuning changes behaviour. RAG changes knowledge.

Use a specialised processor such as invoice or form parser, with human review for low confidence.

Domain-tuned parsing beats raw OCR for structured fields.

Ingest the catalogue and user events and use Vertex AI Search for commerce.

It is the industry-specific product, built from events and catalogue.

Provide time column, target, series ID, horizon and mark covariates available at forecast time.

Correct column roles stop leakage from future data.

Embed the transforms in the model or read shared feature definitions.

Separate reimplementation causes training-serving skew.

Define feature groups in BigQuery and serve through an online store.

Offline joins are point-in-time correct and the online store answers by entity ID.

Run an Apache Beam pipeline, which can also be the TFX runner.

Serverless, autoscaling, same code for batch and stream.

Inspect and de-identify before it reaches notebooks or training, and add policy tags and VPC Service Controls.

Models can memorise records, so scrub early and restrict access.

Use a dedicated service account, no external IP, CMEK and VPC-SC.

Workbench is the controllable option. Colab Enterprise is the serverless one.

Log parameters and metrics per run and compare in the console.

TensorBoard shows training curves, Experiments compares runs.

Use autorater metrics, pairwise comparison and a golden dataset, with human review.

Computation metrics need references, which open-ended output lacks.

Checkpoint to Cloud Storage and restart on preemption.

Spot is cheap but can be reclaimed.

Enable Reduction Server on the Vertex AI worker pools.

It reduces all-reduce bandwidth for data-parallel GPU jobs.

Split the model across devices (FSDP, tensor parallel) or use a larger accelerator.

Data parallelism needs the full model on each worker.

Define a double parameter with log scaling in a Vertex AI tuning job.

Linear scaling wastes trials on large values.

Clip gradients, normalise inputs and look for inf or NaN values.

Exploding gradients or bad inputs cause it.

Remove features unknown at prediction time and fit transforms on the training split only.

Leakage inflates validation results.

Use sharded files, parallel reads, prefetch and the profiler.

The accelerator is starved of data.

Run Vertex batch prediction or BigQuery ML nightly and store results.

No always-on endpoint cost.

Deploy to a Vertex AI endpoint with warm minimum replicas.

Batch cannot meet request-time latency.

Deploy both versions to one endpoint and shift a small percentage.

Rollback is a traffic change.

Mirror requests, log the candidate's output and compare offline.

Users never see the candidate's response.

Create a private endpoint reachable only through a private IP, plus VPC Service Controls.

Traffic stays on Google's private network.

Build an image exposing health and predict routes, or use a custom prediction routine.

Prebuilt images cover standard frameworks only.

Reduce precision or train a smaller student, and compile with TensorRT or XLA.

Each shrinks compute and memory with limited accuracy loss.

Read precomputed features by entity ID and merge with the request.

Computing aggregates per request is too slow.

Define components in KFP or TFX and run on a schedule.

Serverless, with metadata and caching built in.

Use Airflow to orchestrate the DAG and launch a Vertex AI Pipeline.

Composer is the general orchestrator.

Run KFP on Kubernetes, with containers and open formats.

Vertex AI Pipelines is Google Cloud only.

Gate registration and deployment on metrics against the champion and slices.

Prevents regressions from automated retraining.

Send a Model Monitoring alert to Pub/Sub and start the pipeline from Cloud Run.

Labels are delayed, so drift is the early signal.

Run tests, build images, compile the pipeline and deploy through environments.

Native CI that triggers from Git.

Trace the model artifact to its execution, dataset and run, with Model Registry links.

Metadata stores artifact and execution relationships.

Compare serving data against the training dataset and fix the preprocessing.

Skew is a pipeline problem, not a retraining problem.

Compare recent serving data with earlier serving data and retrain.

Drift needs no training baseline.

Enable Explainable AI attributions in Model Monitoring.

Detects the model relying on different features.

Compute fairness metrics per subgroup with TFMA or Model Evaluation slices.

Aggregates hide failing segments.

Configure Sampled Shapley attributions at model upload.

It works on non-differentiable models.

Filter inputs and outputs, use safety filters, and give agent tools least privilege.

Retrieved text can carry hidden instructions.

Increase max replicas or quota, or use provisioned throughput.

429 means overload or throttling, not a model bug.