#Intro

Will Sewell
Platform engineer at Pusher
Slides
Will talk about garbage collection

Pusher とは

Hosted web socket API
Users publish using http
Users subscribe via web sockets
- Recieve messages via this
Based on Redis hub/sub

The problem

No state
Brief Loss of connection => Will never recieve message
Need to keep cache / message queue
Need to handle >10,000 / second to be affordable
Latency must be <50ms (Publish => Subs)

Why Haskell?

Made before speaker joined company
Legacy written in Ruby
After codebase grows, start to crave haskell
Like simple languages with hard principles
Fast

Pros / Cons of Haskell

Just speaker's opinion

Pros

Types, good for correctness as well as design
Concurrency Model
- Green threads, STM
Great tooling
- Threadscope
- Heap visualization
- Stack
Hiring
- Lots of high quality candidates

Cons

No standard way of solving problems
Lots of competing solutions
- Strings vs Text
- Pipes v Conduit
Error messages
- Abstractions leak
- Need to understand lots of underlying concepts
Tooling limitations
- Some are good, but limited options and often clunky
- Not living up to potential, tools could be so much better

The Message Store

Problems with Garbage Collection

Messages constantly inserted and removed from store
Garbage collection blocking application from running
Max pause time proportional to size of working set

GHC GC Design

High rate of un-referencing
Most objects die young
Generational
- Heap split in generations
- Long-term objects exist in 'old generation' heap
- Checked infrequently
'Stop The World'
- Application blocked while heap being checked

Worst Case

Message Store has long and equal life times
All objects in old generation

Solutions

Write storage in different language
Write program in different language
- Re-wrote program in Go
- Production raft libraries
- GC Specifically low latency