hydro-serving-cli

CLI tool for hydro-serving.

Installation

pip install hs

Usage

hs cluster

Since 0.1.0 the tool uses configurable endpoints called cluster: They can be managed with:

• hs cluster - shows default cluster
• hs cluster add --name local --server http://localhost - creates entry with name local and address http://localhost.
• hs cluster use local - uses the previously created local cluster as default
• hs cluster rm local - removes local cluster entry

The default cluster is used as endpoint for all API calls made from the CLI tool.

hs upload

When you use hs upload, the tool looks for serving.yaml file in current dir. serving.yaml defines model metadata and contract

kind: Model
name: "example_model"
install-command: "pip install requirements.txt"  # if you need to run additional command before deployment
runtime: "hydrosphere/serving-runtime-python:3.6-latest"
payload:
  - "saved_model.pb"
  - "variables/"

monitoring:
  - name: ks
    kind: KSMetricSpec
    config:
      input: client_profile_42
    with-health: true
  - name: gan
    kind: GANMetricSpec
    with-health: true
    config:
      input: feature
      application: claims-gan-app
  - name: autoencoder
    kind: AEMetricSpec
    with-health: true
    config:
      application: claims-autoencoder-app
      input: feature
      threshold: 69
  
contract:
  name: detect  # the name of signature
  inputs:  # signature input fields
    image_b64:
      type: string
  outputs:  # signature output fields
    scores:
      shape: [-1]
      type: double
    classes:
      shape: [-1]
      type: string

hs apply

The command takes -f path parameter to yaml file or directory containing yaml files. Then, it will read the yaml documents and apply them sequentially.

These files can contain definition of a resource defined below:

Model

The model definition is the same as in serving.yaml file.

HostSelector

Example of a selector definition:

kind: HostSelector
name: xeon-cpu
node_selector: "/* INSTANCE SELECTOR */"

Note that selector is a string that defines platform specific filter on instances.

Application

For the sake of simplicity CLI provides simplified structures for major use cases:

• Single model app:

kind: Application
name: demo-app
singular:
  model: demo_model:2

• Pipeline app

kind: Application
name: demo-pipeline-app

pipeline:
  - - model: claims-preprocessing:1
  - - model: claims-model:1
      weight: 80
    - model: claims-model:2
      weight: 20