h2o-mlops-scoring-client

H2O MLOps Scoring Client

A Python client library to simplify robust mini-batch scoring against an H2O MLOps scoring endpoint. It can run on your local PC, a stand alone server, Databricks, or a Spark 3 cluster.

Scoring Pandas data frames is as easy as:

pip install h2o-mlops-scoring-client

import h2o_mlops_scoring_client


scores_df = h2o_mlops_scoring_client.score_data_frame(
    mlops_endpoint_url="https://.../model/score",
    id_column="ID",
    data_frame=df,
)

Scoring from a source to a sink is also possible through pyspark:

pip install h2o-mlops-scoring-client[PYSPARK]

import h2o_mlops_scoring_client


h2o_mlops_scoring_client.score_source_sink(
    mlops_endpoint_url="https://.../model/score",
    id_column="ID",
    source_data="s3a://...",
    source_format=h2o_mlops_scoring_client.Format.CSV,
    sink_location="s3a://...",
    sink_format=h2o_mlops_scoring_client.Format.PARQUET,
    sink_write_mode=h2o_mlops_scoring_client.WriteMode.OVERWRITE
)

Installation

Requirements

• Linux or Mac OS (Windows is not supported)
• Java (only required for pyspark installs)
• Python 3.8 or greater

Install from PyPI

pip install h2o-mlops-scoring-client

pyspark is no longer included in a default install. To include pyspark:

pip install h2o-mlops-scoring-client[PYSPARK]

FAQ

When should I use the MLOps Scoring Client?

Use when batch scoring processing (authenticating and connecting to source or sink, file/data processing or conversions, etc.) can happen external to H2O AI Cloud but you want to stay within the H2O MLOps workflow (projects, scoring, registry, monitoring, etc.).

Where does scoring take place?

As the batch scoring processing occurs, the data is sent to an H2O MLOps deployment for scoring. The scores are then returned for the batch scoring processing to complete.

What Source/Sinks are supported?

The MLOps scoring client can support many source/sinks, including:

• ADLS Gen 2
• Databases with a JDBC driver
• Local file system
• GBQ
• S3
• Snowflake

What file types are supported?

The MLOps scoring client can read and write:

• CSV
• Parquet
• ORC
• BigQuery tables
• JDBC queries
• JDBC tables
• Snowflake queries
• Snowflake tables

If there's a file type you would like to see supported, please let us know.

I want model monitoring for batch scoring, can I do that?

Yes. The MLOps Scoring Client uses MLOps scoring endpoints which are automatically monitored.

Is a Spark installation required?

No. If you're scoring Pandas data frames, then no extra Spark install or configuration is needed. If you want to connect to an external source or sink, you'll need to install pyspark and do a small amount of configuration.