Nuclio functions

Nuclio - High performance serverless for data processing and ML

Nuclio Overview

nuclio is a high performance serverless platform which runs over docker or kubernetesand automate the development, operation, and scaling of code (written in 8 supported languages).Nuclio is focused on data analytics and ML workloads, it provides extreme performance and parallelism, supports stateful and data intensiveworkloads, GPU resource optimization, check-pointing, and 14 native triggers/streaming protocols out of the box including HTTP, Cron, batch, Kafka, Kinesis,Google pub/sub, Azure event-hub, MQTT, etc. additional triggers can be added dynamically (e.g. Twitter feed).

nuclio can run in the cloud as a managed offering, or on any Kubernetes cluster (cloud, on-prem, or edge)read more about nuclio …

Using Nuclio In Data Science Pipelines

Nuclio functions can be used in the following ML pipline tasks:

  • Data collectors, ETL, stream processing
  • Data preparation and analysis
  • Hyper parameter model training
  • Real-time model serving
  • Feature vector assembly (real-time data preparation)

Containerized functions (+ dependent files and spec) can be created directly from a Jupyter Notebookusing %nuclio magic commands or SDK API calls (see nuclio-jupyter),or they can be built/deployed using KubeFlow Pipeline (see: nuclio pipeline components)e.g. if we want to deploy/update Inference functions right after we update an ML model.

Installing Nuclio over Kubernetes

Nuclio Git repo contain detailed documentation on the installation and usage.can also follow this interactive tutorial.

The simplest way to install is using Helm, assuming you deployed Helm on your cluster, type the following commands:

  1. helm repo add nuclio https://nuclio.github.io/nuclio/charts
  2. kubectl create ns nuclio
  3. helm install nuclio/nuclio --name=nuclio --namespace=nuclio --set dashboard.nodePort=31000
  5. kubectl -n nuclio get all

Browse to the dashboard URL, you can create, test, and manage functions using a visual editor.

Note: you can change the NodePort number or skip that option for in-cluster use.

Writing and Deploying a Simple Function

The simplest way to write a nuclio function is from within Jupyter.the entire Notebook, portions of it, or code files can be turned into functions in a single magic/SDK command.see the SDK for detailed documentation.

The full notebook with the example below can be found here

before you begin install the latest nuclio-jupyter package:

  1. pip install --upgrade nuclio-jupyter

We write and test our code inside a notebook like any other data science code.We add some %nuclio magic commands to describe additional configurations such as which packages to install,CPU/Mem/GPU resources, how the code will get triggered (http, cron, stream), environment variables,additional files we want to bundle (e.g. ML model, libraries), versioning, etc.

First we need to import nuclio package (we add an ignore comment so this line wont be compiled later):

  1. # nuclio: ignore
  2. import nuclio

We add function spec, environment, configuration details using magic commands:

  1. %nuclio cmd pip install textblob
  2. %nuclio env TO_LANG=fr
  3. %nuclio config spec.build.baseImage = "python:3.6-jessie"

and we write our code as usual, just make sure we have a handler function whichis invoked to initiate our run. The function accepts a context and an event, e.g.: def handler(context, event)

Function code

the following example show accepting text and doing NLP processing (correction, translation, sentiments):

  1. from textblob import TextBlob
  2. import os
  4. def handler(context, event):
  5.     context.logger.info('This is an NLP example! ')
  7.     # process and correct the text
  8.     blob = TextBlob(str(event.body.decode('utf-8')))
  9.     corrected = blob.correct()
  11.     # debug print the text before and after correction
  12.     context.logger.info_with("Corrected text", corrected=str(corrected), orig=str(blob))
  14.     # calculate sentiments
  15.     context.logger.info_with("Sentiment",
  16.                              polarity=str(corrected.sentiment.polarity),
  17.                              subjectivity=str(corrected.sentiment.subjectivity))
  19.     # read target language from environment and return translated text
  20.     lang = os.getenv('TO_LANG','fr')
  21.     return str(corrected.translate(to=lang))

Now we can test the function using a built-in function context and examine its output

  1. # nuclio: ignore
  2. event = nuclio.Event(body=b'good morninng')
  3. handler(context, event)

Finally we deploy our function using the magic commands, SDK, or KubeFlow Pipeline.we can simply write and run the following command a cell:

%nuclio deploy -n nlp -p ai -d <nuclio-dashboard-url>

if we want more control we can use the SDK:

  1. # nuclio: ignore
  2. # deploy the notebook code with extra configuration (env vars, config, etc.)
  3. spec = nuclio.ConfigSpec(config={'spec.maxReplicas': 2}, env={'EXTRA_VAR': 'something'})
  4. addr = nuclio.deploy_file(name='nlp',project='ai',verbose=True, spec=spec, 
  5.                           tag='v1.1', dashboard_url='<dashboard-url>')
  7. # invoke the generated function 
  8. resp = requests.get('http://' + addr)
  9. print(resp.text)

We can also deploy our function directly from Git:

  1. addr = nuclio.deploy_file('git://github.com/nuclio/nuclio#master:/hack/examples/python/helloworld',
  2.                           name='hw', project='myproj', dashboard_url='<dashboard-url>')
  3. resp = requests.get('http://' + addr)
  4. print(resp.text)

Using Nuclio with KubeFlow Pipelines

We can deploy and test functions as part of a KubeFlow pipeline step.after installing nuclio in your cluster (see instructions above), you can run the following pipeline:

  1. import kfp
  2. from kfp import dsl
  4. # load nuclio kubeflow components
  5. nuclio_deploy = kfp.components.load_component(url='https://raw.githubusercontent.com/kubeflow/pipelines/master/components/nuclio/deploy/component.yaml')
  6. nuclio_invoke = kfp.components.load_component(url='https://raw.githubusercontent.com/kubeflow/pipelines/master/components/nuclio/invoker/component.yaml')
  8. @dsl.pipeline(
  9.     name='Nuclio deploy and invoke demo',
  10.     description='Nuclio demo, build/deploy a function from notebook + test the function rest endpoint'
  11. )
  12. def nuc_pipeline(
  13.    txt='good morningf',
  14. ):
  15.     nb_path = 'https://raw.githubusercontent.com/nuclio/nuclio-jupyter/master/docs/nlp-example.ipynb'
  16.     dashboard='http://nuclio-dashboard.nuclio.svc:8070'
  18.     # build the function image & CRD from a notebook file (in the above URL)
  19.     build = nuclio_deploy(url=nb_path, name='myfunc', project='myproj', tag='0.11', dashboard=dashboard)
  21.     # test the function with real data (function URL is taken from the build output)
  22.     test = nuclio_invoke(build.output, txt)

the code above assumes nuclio was deployed into the nuclio namespace on the same cluster, when using a remote cluster or a different namespace you just need to change the dashboard URL.

See nuclio pipline components (allowing to deploy, delete, or invoke functions)

Note: Nuclio is not limited to Python, see this example showing how we create a simple Bash function from a Notebook, e.g. we can create Go functions if we need performance/concurrency for our inference.

Nuclio function examples

Some useful function example Notebooks: