dask/dask

Please use Stack Overflow with the #dask tag for usage questions and github issues for bug reports

Bradley McElroy

@limx0

I'm interested in gathering some worker resource stats for some dask graph runs

Matthew Rocklin

@mrocklin

All dask workers run the system monitor internally

Bradley McElroy

@limx0

Okay good to know, do the kubes workers record their usage like the local cluster? I'm actually trying to do a range query after some computation, is this also available on the kubes cluster?

Matthew Rocklin

@mrocklin

(FYI, I don't answer questions here as a rule. if you're looking for help from me personally you'll have to use stack overflow or github)

Bradley McElroy

@limx0

understood, thanks @mrocklin - dask/dask-kubernetes#180

Matthew Rocklin

@mrocklin

I appreciate it!

Adam Thornton

@athornton

I just opened this issue: dask/dask-kubernetes#181 but figured someone here might know. What changed in dask-kubernetes with respect to RBAC after 0.9.1 ? I get a 403 Forbidden:

User \"system:serviceaccount:nublado-athornton:dask\" cannot get resource \"pods\" in API group \"\" in the namespace \"nublado-athornton\""

but I have what look like the right rules in my role:

rules:
- apiGroups:
  - ""
  resources:
  - pods
  verbs:
  - list
  - create
  - delete

Eric Ma

@ericmjl

@TomAugspurger I hope you're doing well. I've got a question for you regarding dask-ml - is there a way to prevent parallel random forest fitting from using >100% CPU? I have set n_jobs=1 in the RandomForestRegressor() constructor, but still end up with some of my dask workers using 2000% CPU, which is looking really weird.

(It also makes me a bad citizen on our compute cluster, haha.)

FWIW, i

I've been using dask as part of an automated ML system - one RF per slice/subset of data in our database - and there's like thousands of slices.

It's super cool that Dask enables me to build this out.

That said, I get a lot of garbage collection errors whenever I'm using parallel RF on this many slices of data - is it because of overwhelming the dask scheduler?

Matthew Rocklin

@mrocklin

@ericmjl it might be better to raise this as an issue on github

That way other people who run into the same issue will be able to do a web search and benefit from the answer

Eric Ma

@ericmjl

OK! Thanks @mrocklin! :D

Scott Sievert

@stsievert

I’m referencing dask-jobqueue in an academic paper. What’s the preferred method of referencing it?

I’m currently planning on linking to the readthedocs homepage.

joeschmid

@joeschmid

@stsievert Try this: https://docs.dask.org/en/latest/cite.html

anakaine

@anakaine

Random question: When createding a new column in a dataframe by returning the results from an np.select operation dask complains that the data coming back is not supported, being of type numpy.ndarray. Doing the same in pandas (same code, just one dataframe is set up in dask, the other in pandas) works fine. Is this as expected?

Scott Sievert

@stsievert

Thanks for that @joeschmid

Simon

@anakaine_twitter

Is there an equivalent of np.select in dask? Checking the docs theres quite a few references to "see also: select", but no section for it.

Simon

@anakaine_twitter

https://stackoverflow.com/questions/58277992/dask-is-there-an-equivalent-of-numpy-select-for-dask

David Hoese

@djhoese

Not sure of a better place for this, but I'm looking for a small example of using fastparquet that shows the advantages of defining row-groups over the default "single" row group. I'm presenting some information on file formats but don't use parquet in my day to day (I really don't even use pandas that often). Anyone have a simple code example?

Uwe L. Korn

@xhochy

@djhoese Here's a very simple example: https://gist.github.com/xhochy/0880fd0bed3a7eaca3c31bf07b6868d9 (although it uses pyarrow)

Martin Durant

@martindurant

correct, whether or not to use row-groups is not specific to the engine; Dask will load in parallel with either and can make (limited) decisions to exclude some row groups based on metadata.

David Hoese

@djhoese

@xhochy Ok thanks. Is there a way to do that query without knowing that row-group 1 is where you want to look?

Does generating a pandas DataFrame (not dask) load the entire parquet file in to memory? Or does it do it lazily?

Martin Durant

@martindurant

That last time I looked, pandas loaded everything. It would be reasonable to implement that iteratively, and fastparquet does have a specific method to do that.

Is there a way to do that query without knowing that row-group 1 is where you want to look

Parquet optionally stores columns max and min values for each row-group, so maybe

David Hoese

@djhoese

Thanks @martindurant

Davis Bennett

@d-v-b

an interesting blog post about designing a threaded scheduler in rust: https://tokio.rs/blog/2019-10-scheduler/

Dean Langsam

@DeanLa

how do I expand a series containing fixed-size list into a data frame?

https://stackoverflow.com/questions/58384320/expand-array-to-columns-in-dask-dataframe

Scott Sievert

@stsievert

@DeanLa https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.explode.html (tried answering on SO)

Dean Langsam

@DeanLa

not what i need.
i need to expand a list to columns (on axis 1), not explode and create new rows

Sergio Meana

@smeana

Hello
Is any build in function to do forward rolling on timeseries index?
I have seen same question asked by @AlbaFS on Aug 04 without reply. Any solution at all? thks

mcguip

@mcguipat

Hi all, I have a problem which requires data to be loaded in several blocks which are then subdivided and operated on. Scatter seems like the most sensible way of accomplishing this; however, if I want loading to be executed on the cluster as well, this requires scatter from within another task. Is there a better way to accomplish this?

Dario Vianello

@dvianello

Hey! I'm struggling a bit with s3 & IAM roles while using dask. We have a set of workers in EC2 with instance profiles authorised to assume a role in a different account and read from a bucket. I've tried to do the assume role beforehand in boto3 and pass the session into dask, but the session object can't be pickled apparently (fails to pickle the lock). Is there a way to pull this off in Dask? Sending creds along to workers isn't the best idea ever and it would be cool if the system was able to do the assume role on the workers before trying to access s3...

Martin Durant

@martindurant

Whatever you do to assume the role, you could execute the same thing on the cluster workers using client.run

Dario Vianello

@dvianello

right!

Dean Langsam

@DeanLa

Running dask-yarn, notebook crashed, I think Cluster instance (scheduler?) is dangling because I can't create a new one. How can I make sure?

xavArtley

@xavArtley

Hello,
I'm using

dask.compute(*delayeds, scheduler='processes', num_workers=4)

to run computations.
I was wondering which function was used to serialized object between processes and if it was possible to change it
Thanks

Fred Massin

@FMassin

Hello,
I would like to use template matching with time-series. The objective is to look for many relatively short 1d pattern in a relatively long time-series. Any suggestion on how to do this in Dask? I mean to have something like https://docs.dask.org/en/latest/dataframe-api.html#dask.dataframe.Series.corr with many other and split_every equal or lower than length of other...
Thanks!

Berk Gercek

@berkgercek

Heya, hate to add to the pile of questions but I'm currently going through dask-tutorial and I am on the weather example in Ch. 2: Arrays. The computation of mean in 500x500 chunks takes 14m31s of time (not wall time), with only 1m50s of that being User time and the rest is sys. Is this a normal result? The wall time for reference is ~37s, and I am using a 12-core processor

Martin Durant

@martindurant

^ many of these sounds like stack-overflow questions or issues (to tutorials in the latter case). We discourage anything more than general conversation here, because we would like solutions to be searchable, so others can benefit too

Fred Massin

@FMassin

Matthew Rocklin

@mrocklin

+1 to what @martindurant said. @xavArtley @FMassin @berkgercek those are all great and valid questions. Most of the Dask maintainers don't watch this channel often. See https://docs.dask.org/en/latest/support.html for suggestions on where to ask for help

xavArtley

@xavArtley

Hello, I post here https://stackoverflow.com/questions/58449867/change-serialisation-protocol-for-multiprocessing-windows

Fred Massin

@FMassin

Mine: https://stackoverflow.com/questions/58439177/template-matching-for-time-series-using-dask

Martin Durant

@martindurant

(it may still take some time for replies, but the audience is much broader)

lszyba1

@lszyba1

Hello,
I'm trying to load a csv file, but there are some inconsistent column types...
So while initially the file looks like bunch of 0, then its an object...
is there a way to:
a) parse through all the csv file 10gb+ and create correct dtype. (I have over 100 columns)
or
b) go through the file and summersize the type counts... for example is there 99,999 rows of int and 1 row of object...and I can just clean that row?

Where communities thrive

People

Repo info

Activity