serverless-presentation/index.rst

188 lines
3.7 KiB
ReStructuredText

=====================================
Serverless real time video processing
=====================================
Powered by the Cloud, AI, Microservices, and other buzzwords
The Serverless hype
===================
What does serverless mean?
--------------------------
.. image:: _static/empty_racks.jpg
.. Advantages of serverless:
- No management
- Faster startup times
- "Infinite scale"
- Business logic only
AWS Compute startup times
-------------------------
.. revealjs_fragments::
* EC2 - minutes
* Fargate - seconds
* Lambda - milliseconds (when warm and cozy)
* https://firecracker-microvm.github.io/
The Lambda lifecycle
--------------------
.. code-block:: python
print('initialized')
def handler(event, context):
print('executed')
return {
'message': 'done'
}
.. What is a cold start and the advantages of a warm lambda
Edifice Corb
============
.. image:: _static/corbusier.jpg
* Named after some weird French dude
* Real time person detection
* User interface through SMS
Rapid iteration
---------------
.. image:: _static/say_agile.jpg
.. Constantly reworking the infrastructure and sometimes even the business logic
Person detection
================
* Inference on video frames is compute intensive
* GPU instances are neither serverless nor cheap
.. image:: _static/god_damn_aws_charges.jpg
:width: 60%
.. Don't use GPU for inference unless absolutely necessary
AWS to the rescue
-----------------
.. image:: _static/behold_elastic_inference.jpg
:width: 40%
Still not serverless
--------------------
.. image:: _static/elastic_inference.png
:width: 40%
The AWS SageMaker of six paths
------------------------------
Managed end-to-end ML service
.. image:: _static/sagemaker_studio.jpg
:width: 80%
Processing pipeline
===================
.. image:: _static/processing_pipeline.png
Triggers and destinations
-------------------------
.. image:: _static/triggers_and_destinations.png
Serverless framework
--------------------
.. image:: _static/serverless.jpg
:width: 30%
True yaml engineering
.. No manual zip files, less architecture management, but less flexible
Triggering events
-----------------
.. code-block:: yaml
functions:
VideoIngress:
handler: aws_lambda.run_detector.handler
events:
- s3:
bucket: video-bucket
event: s3:ObjectCreated:*
rules:
- suffix: .mkv
Other serverless goodies
------------------------
* Automatic IAM roles
* Easy additional permissions
* Lots of functionality for little code
.. show serverless.yaml
Other serverless badies
-----------------------
* CloudFormation bad, Terraform good
* Multiple environments are harder to implement
* Hardcoding and code duplication
.. show serverless.yaml
Messaging pipeline
==================
.. image:: _static/messaging_pipeline.png
:width: 40%
.. Current architecture. Will be more complex when we add alarms. Talk about sync/async lambda calls
Y tho?
------
.. image:: _static/why_few_lambda.jpg
.. Why microservices. Easier parallel development. Easier testing
More triggers and destinations
------------------------------
.. code-block:: yaml
CleaningReportGenerator:
handler: aws_lambda.request_cleaning.handler
events:
- eventBridge:
pattern:
detail:
responsePayload:
command_type:
- request_cleaning
destinations:
onSuccess: arn:aws:events:::event-bus/default
EventBridge - Central point of success
--------------------------------------
* Arbitrary event routing and scheduling
* All services produce events
* Cheap af (probably even free)