GCP / BigQuery FOR AI ENGINEERS — AWS-to-GCP Translation Crash Course The Home Depot · Principal AI Engineer · Round 1 · 2026-05-01

Mental Model

GCP is Google's cloud. Home Depot has run on GCP exclusively for 10+ years (600+ projects). BigQuery is their analytical data warehouse — 15+ petabytes, migrated from 450TB on-prem. BigQuery is not a real-time database — it's for analytics, model training, and historical pattern mining. Real-time inference runs on GKE with Pub/Sub as the event bus.


For Bishop: Architecture patterns are identical to AWS. The vendor labels are different. BigQuery = Redshift (serverless). Vertex AI = SageMaker. GKE = EKS. Pub/Sub = Kinesis. The API surface is the learning curve — not the architecture.

Home Depot's Actual Stack CONFIRMED

LayerHD Tool
CloudGCP only (10-yr partnership, Jan 2026 expanded)
Data WarehouseBigQuery (15+ PB, serverless)
ML PlatformVertex AI (training, registry, serving)
LLM / Gen AIGemini Enterprise (associates + Magic Apron)
StreamingCloud Pub/Sub + Dataflow
ContainersGKE (Kubernetes, Docker)
Legacy Spark/HadoopDataproc (managed)
BackendJava Spring Boot (primary), Python (data science)
EnterpriseSAP S/4HANA on GCP (migrated 2017–2021)
RoboticsSIMPL Automation (G2P, acquired recently)

AWS → GCP Service Translation

AWS (Bishop knows)GCP (HD uses)Notes
S3Cloud Storage (GCS)Object storage, identical pattern
RedshiftBigQueryBigQuery is serverless — no cluster to manage
SageMakerVertex AITraining jobs, model registry, endpoints, pipelines
KinesisCloud Pub/SubStreaming event bus — real-time WES events here
Kinesis AnalyticsDataflowStreaming ETL / stream processing (Apache Beam)
EKSGKEManaged Kubernetes — where real-time AI inference runs
LambdaCloud Functions / Cloud RunServerless compute
EMR / SparkDataprocNaaga's Hadoop background lives here
EC2Compute EngineVMs
RDSCloud SQL / SpannerManaged relational DB
CloudWatchCloud Monitoring / LoggingObservability
IAMCloud IAMSame concept, different syntax

BigQuery Key Facts

  • Serverless: No cluster to provision. Query petabytes in seconds, pay per TB scanned.
  • Partitioned tables: Partition warehouse event data by timestamp for efficient queries.
  • BigQuery ML: Run logistic regression, boosted trees, time series — directly in SQL.
  • Streaming inserts: Real-time data streams directly into BigQuery tables via API.
  • NOT for real-time inference: Latency is seconds-to-minutes. Hot-path WES decisions run on GKE services, not BigQuery.

How GCP Maps to WES Architecture

Cloud Pub/Sub — Event Bus
Real-time warehouse events (robot state, station load, conveyor status, order updates). Sub-second latency. The nervous system of the WES.
GKE — Real-Time Inference Layer
Where the carton streaming scoring model runs. Python microservice containerized and deployed on GKE. Receives events from Pub/Sub, returns decisions. This is the hot path.
BigQuery — Historical Analytics + Training Data
Where HD's 15+ PB lives. Use for: feature engineering (historical pick times, congestion patterns), model training datasets, A/B experiment analysis, cut-time compliance reporting.
Vertex AI — Model Lifecycle
Model training jobs, experiment tracking, model registry, serving endpoints. Equivalent to SageMaker Pipelines + Model Registry. Bishop frames: "I'd use Vertex AI the same way I've used MLflow and SageMaker for model lifecycle management."
Dataflow — Stream Processing
For enriching real-time events with contextual features before scoring. Joins live Pub/Sub stream with BigQuery lookup tables (inventory, station capacity, order priorities).

Interview Language (Use Verbatim)

"BigQuery is your analytical foundation — I'd use it for historical pattern mining and training data, not real-time inference."
"The real-time decision layer runs on GKE with Pub/Sub as the event bus — same architecture as Kinesis on AWS."
"Vertex AI is my model registry and serving layer — I'd use it the same way I've used SageMaker for model lifecycle management."
"I'm AWS-primary by experience, but the architectural patterns are identical — the GCP-specific APIs are a sprint-one learning curve, not a design blocker."
"I'd keep the scoring model hot path on GKE, not BigQuery — BigQuery is seconds-to-minutes latency, and carton streaming needs sub-second."