What’s generally asked in the ML infra system design interview as opposed product(full stack) or platform? I’ve an interview coming so any pointers will be helpful.#systemdesign #interview
Following
F
It's so disappointing that ML system design is still a gray area, there are very few contents available and people can't answer clearly what they faced or what they ask.
Everyday there’s a post asking same exact question. Seems like a good opportunity to make a course like grokking
Yep. Can't find anything specific on the internet.
Ask your recruiter for prep advice
Hey! I am also giving machine learning system design interviews. Pm me we can mock and give each other feedback.
When's your interview, Apple. I have something coming up
Here are a few, from my last FB ML infra round: 1) design Feed relevance. a) what kind of basic inputs do you need (1st degree connection graph, will need infra to get some kind of similarity metric put on it to differentiate between acquaintances and close friends; will need the text of the posts, metadata of whether its img/video/etc; will need the “like” stream as well as other interactions (clicks, comments, seconds spent watching the video, hover time, etc). b) what kind of data stores to handle both serving throughput / latency, as well as offline compute to do the relevance side. c) what kind of real-time inputs are possible with the infra you are designing so far (trending data? in-session interactions?) 2) design a model store: how does it handle model versioning, backwards compatibility, a/b testing, easy prototyping? 3) you’re on a team which does ML totally offline: trains SparkML models and computes results in hourly batch jobs. Design how to migrate this to an inference during the request path system. This was all for E6/E7 loop, but similar questions with lower expectations for E5 should hold.
following
How do add versioning for the model store? Ab testing isnt usually part of the model store? Please give me some feedback.
Following