I85PREP
⌘ K
H2O.ai Certified Machine Learning study materials
H2O.ai Certified Machine Learning study resources

H2O.ai Certified Machine Learning: H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

H2O.ai Certified Machine Learning · Work through your exam, one topic at a time.

All guide sections

H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

What this domain tests

The H2O Platform Architecture and Deployment domain tests how H2O-3 and Driverless AI are structured, how clients talk to them, and how models leave the training environment. Expect questions on cluster components, REST communication, MOJO/POJO artifacts, and the difference between an H2O-3 AutoML run and a Driverless AI experiment.

Core rules

  • H2O-3 is a distributed, in-memory analytics engine. A cluster has nodes, a distributed key/value store (DKV), and a REST API. Python, R, Scala, Java, and Flow are clients; they do not hold the training data.
  • h2o.init() connects to or starts a cluster. Size memory for the data plus intermediates, not just the raw file. H2O-3 can run on a laptop, Kubernetes, Hadoop/YARN, or Sparkling Water; Spark is optional, not required.
  • Data is parsed into an H2OFrame, which is distributed across the cluster. Losing a node can lose data; H2O-3 is not a fault-tolerant database.
  • Driverless AI is a separate single-node platform, often GPU-accelerated, built around experiments. It performs automatic feature engineering, model selection, and interpretability, and produces a Python scoring pipeline and a MOJO.
  • H2O-3 models can be exported as POJO or MOJO. POJO is Java source; MOJO is a compact binary artifact. For external scoring, MOJO plus the h2o-genmodel.jar runtime is the usual choice.
  • Driverless AI scoring pipelines include feature engineering; a bare MOJO contains the model, not the full recipe pipeline.

Worked example

A Java microservice needs sub-50 ms credit decisions and cannot call an H2O cluster. Train a GBM in H2O-3, export the MOJO with h2o.download_mojo(get_genmodel_jar = TRUE), and embed it with the MOJO runtime in the service. For a weekly batch job in Python that must reproduce Driverless AI feature engineering, use the exported Driverless AI scoring pipeline instead. The reasoning is deployment latency plus preprocessing ownership.

Traps

  • Assuming H2O-3 requires Spark or Hadoop.
  • Confusing a Driverless AI experiment with h2o.automl().
  • Thinking a MOJO needs a live H2O cluster to score.
  • Using Flow as a production scoring endpoint.
  • Forgetting that Driverless AI pipelines and H2O-3 MOJOs package different amounts of preprocessing.