[QCon 2019] State of Serverless Computing

Chenggang Wu – Berkley

For other QCon blog posts, see QCon live blog table of contents

General

  • 2000 students in intro class at Berkley last year. Put them in concert hall
  • How make them more productive so can work on cheap machines like Chromebooks
  • Hosted infrastructure/cloud
  • Companies dedicated to making serverless easy

FaaS (Function as a Service)

  • Run code and pay for only what use
  • ex: AWS Lambda
  • Optimized for simplicity
  • Good at
    • embarassingly parallel tasks – video processing, translation
    • workflow orchestration
  • Limitations
    • Limited execution time – AWS increased from 1 minute to 5 minutes and then again to 15 minutes
    • No inbound network connection
    • IO is a bottleneck
    • Doesn’t support specialized hardware
    • Not designed for functional programming because real applications share state. Also, no natural way to chaining/composing functions. Using hacks to get around that introduces latency. AWS Step Functions even slower than hacking with S3 or Dynamo.

BaaS (Back End as a Service)

  • ex: AWS Athena
  • more specific

Storage

  • Want both autoscaling and low latency
  • In practice, improving one makes the other worse.
  • Autoscaling systems also provide poor consistency guarantees

Berkley’s work

  • Platform for serverless computing
  • working on supporting state
  • Anna – data store used in place of S3. Chenggang worked on for PHD
  • Fluent – FaaS over Anna. Working on new name since already a project named Fluent. Use Anna for both storage and communication reduces crossing network boundary.
  • Cache data to increase performance
  • Lattice – data structure that accepts incoming data. Preserves associate, community and idempotent properties.
  • Casual consistency – strongest consistency level that doesn’t require coordination

My impressions

Good end to the sessions for the day. The material was well laid out and easy to understand. The snails were cute when highlighting slowness! I like that most of it wasn’t specific to the research project at Berkley. The Berkley work was interesting, but less applicable.