I'm a mid-level data scientist, 5 years in, mostly doing experimentation and modeling work but I've been leaning heavier into deployment and MLOps over the last year. I want to make the full jump to ML engineer.
My concern: leveling. Every DS-to-MLE thread I find is from 2021 and the market is different now. Are companies taking DS experience at face value when leveling for MLE roles, or are they treating it like a fresh hire?
Specific questions: Do they count my years of experience or run me through the same L3/L4 pipeline as a new grad? How much of the coding bar is actually different from what I already do? Is it worth positioning at the intersection ("ML Platform DS") vs. committing to the full MLE label?
Not fishing for a pep talk, actually want the data.
4 replies
ml_mike
made this switch 3 years ago, here's what actually happened. most companies leveled me one notch down: i came in as senior DS, got offers at mid-level MLE. one company leveled me equivalently because i had production deployment work to show. the coding bar is real: you need to be solid on systems design basics, not just model architecture. if your current work is jupyter-only that's the gap to close first.
ds_dmitri
this is exactly what i needed. one level down is what i assumed, good to have a sanity check. i've been doing actual deployments but on a small team so the scale signals aren't great. will focus on that angle.
backend_bekah
from the other side: backend eng who interviewed a lot of DS-to-MLE folks. the thing that actually differentiates them is whether they've thought about failure modes in production. anyone can explain a model. can you explain what happens when the feature store is stale, or when there's training-serving skew you didn't catch? that's the MLE question underneath the MLE question.
numbers_only
small n but: 4 DS-to-MLE offers i know personally, 2024-2025. 3 of them came in one level down. 1 came in equivalent because they had shipped something that handled 10k+ req/sec. the threshold seems to be production load you can actually cite.