https://flyte.org logo
#ask-the-community
Title
# ask-the-community
s

seunggs

06/17/2023, 2:48 AM
Hi, this may be a newbie question, but should I be using burstable instances for flyte workflows or should I avoid them? Since ML/AI workflows typically run for a short duration, I’m wondering if I should avoid burstable instances?
k

Ketan (kumare3)

06/17/2023, 3:38 AM
Ml ai workflows actually can run for days right
s

seunggs

06/17/2023, 11:13 PM
Right - but it runs at full capacity for the whole time so I figured maybe burstable instances are not well suited? Correct me if I’m wrong
j

jeev

06/18/2023, 4:11 PM
no harm in using burstable probably. it’ll just run slower than you might be comfortable with. could consider using spot instances with checkpointing for likely better savings.
s

seunggs

06/18/2023, 4:11 PM
Right that makes sense - thanks
Actually it looks like that might not be the case - see this post if you’re interested: https://blog.coiled.io/blog/burstable-vs-nonburstable.html
“Coiled runs a suite of Python/Dask benchmarks on AWS. This collection of ~100 benchmarks cover common data science, data engineering, and machine learning workloads. A typical run takes around 150 instance-hours (though as we’ll see, this depends on the choice of instance type). When we switched from burstable to non-burstable instance types, we saw a significant reduction in cost. For instance, comparing two individual runs, we saw costs go from $7.13 (burstable) to $3.89 (non-burstable).”
j

jeev

06/18/2023, 4:52 PM
seems reasonable
but it’s probably workload-specific. spot + check pointing will still give you good savings though, assuming you can do it correctly
s

seunggs

06/18/2023, 4:53 PM
Right
k

Ketan (kumare3)

06/18/2023, 4:56 PM
Dask and Flyte are very different- in terms of how they use the machines. But ya if we want max compute performance it might help for sustained compute requirements
s

seunggs

06/18/2023, 4:57 PM
Are there any benchmarks from Flyte?
k

Ketan (kumare3)

06/18/2023, 4:57 PM
For Compute - nope
We are an orchestrator so we benchmark number of tasks and workflows
s

seunggs

06/18/2023, 4:58 PM
Right - so given a similar data/ML workloads, wouldn’t dask’s benchmark still be roughly same on Flyte?
Or am I missing something?
k

Ketan (kumare3)

06/18/2023, 5:08 PM
Might be, have not read it all