Skip to content

Latest commit

 

History

History
69 lines (50 loc) · 3.71 KB

README.md

File metadata and controls

69 lines (50 loc) · 3.71 KB
Raw

Mister - Distributed MapReduce Framework

Repository for the Kubernetes runner and Library for MapReduce Mister Apps

Mister is a study project and as such is not intended to be used in real-world scenarios (see: Limitations)

This repository provides the module: fregoli.dev/mister which exports the package: mister

A Mister cluster depends on the 3 Go apps residing in cmd/:

  • fetch creates the folder structure in the Persistent Volume Claim and downloads the text files of the books used as input for the MapReduce Jobs.
  • coordinator manages the state of the cluster, hosts the RPC endpoint, spins up the worker Pods and orchestates the MapReduce Jobs.
  • dashboard provides the Web interface and the API endpoint.

The mister package is imported by coordinator and dashboard, as well as the standalone Mister Apps.

Writing a Mister App is straightforward:

  • import fregoli.dev/mister
  • create your app, a struct that implements the mister.App interface: 
    type App interface {
    Map(filename string, contents string) []KeyValue
    Reduce(key string, values []string) string
}
  • call mister.RegisterApp(yourApp)

Example of a Wordcount App run with 4 mapper and 2 reducer workers:

mister-demo-wordcount.mov

Example of a Indexer App run with 5 mapper and 3 reducer workers:

mister-demo-indexer.mov

Concepts explored in this study project

  • exploring Go modules systems by:
    • setting up a vanity url
    • creating a Library module that exposes a simple API
    • creating Applications that import the Library module and elegantly take advantage of its Interfaces and API
  • using the k8s.io/client-go package to manage a cluster programmatically:
    • listing pods
    • spawning jobs
    • extracting logs
  • Learning the basic Kubernetes terminology, components and tools, building a cluster and exposing a service from a YAML file
  • writing Go templates with html/template
  • using Go http to serve the Dashboard and the JSON API
  • exploring different concurrency patterns in Go; throughout the history of the project sample implementations have used various facilities such as the sync primitives (locks, read locks, waitgroups), the atomic package, buffered and unbuffered channels

Limitations

In order to manage the scope of the project, some architectural decisions are suboptimal or resemble antipatterns:

  • Inherently runs on a single node
  • Most state is in-memory, some is derived from filesystem
  • The k8s client is using the default service account and in-cluster config
  • RPC endpoints are not secured
  • shell scripts expect mister and its apps to live under the ~/code folder
  • shell scripts always build and publish the container images to the local minikube registry with the fixed version tag 1
  • No way to specify or provide different input files directly from the dashboard, no way to seamlessly make a new Mister App available to the cluster

Takeaways and Reflections

This project was conceived as a way to apply what I have learned by solving the MapReduce lab assignment from the MIT 6.5840 Distributed Systems class in a more real-world scenario, using industry-standard tools.

The initial plan turned out to be more of an excuse to make the first steps in the Kubernetes domain, which ultimately worked, but it was observed that some of the original problems didn't find a 1-to-1 mapping in the new ecosystem and were therefore set aside, partly or wholesale.