Please use Stack Overflow with the #dask tag for usage questions and github issues for bug reports
Hello everyone. Just wanted to know whether I can perform Dask array computations inside a delayed function, something like this.
def dask_delayed_example(X):
mean = X.mean(axis=0)
std = X.std(axis=0)
return X - mean
X = df.to_dask_array(lengths=True)
res = dask_delayed_example(X)Is Dask delayed only meant for parallelizing computation or can it also be used to perform computation on larger-than-memory datasets in a distributed manner?
Hi @varunbpatil, I'm far from being an expert on dask but I think I have a similar use case. And I asked about it a couple days ago in here. @bnaul pointed me at this recipe https://distributed.dask.org/en/latest/task-launch.html - might be what you're after.
I also recall reading in the docs that nesting delays that return delayed is a antipattern of sorts. So I think you want that task scheduling other tasks approach
Hi there ! I'm trying to use dask-ssh as follows:
% dask-ssh --scheduler master1 node1
---------------------------------------------------------------
Dask.distributed v1.27.0
Worker nodes:
0: node1
scheduler node: master1:8786
---------------------------------------------------------------
[ scheduler master1:8786 ] : /home/applis/anaconda/envs/py3v19.04/bin/python -m distributed.cli.dask_scheduler --port 8786
[ worker node1 ] : /home/applis/anaconda/envs/py3v19.04/bin/python -m None master1:8786 --nthreads 0 --host node1 --memory-limit autobut I cannot connect a client to the scheduler and logging does not provide more information. Client('master1:8786') fails with OSError: Timed out trying to connect... and I can not access the web ui at master1:8787. However, running dask-scheduler and dask-workerwork fine. Any suggestion ?
i'm asking because I noticed that using
da.from_array(large_numpy_array) was blowing up workers, as (counterintuitively) the large_numpy_array is not actually partitioned along the expected chunks. Each worker appears to get a full copy of large_numpy_array regardless of the chunking.
i'm looking into the relevant code to see if there's a simple fix
No, the workers do not get a graph at all, they get tasks with inputs. Depending on the chunking scheme and the operation you are doing, tasks may require many pieces of the overall array. However, if you are doing from_array, it would seem it fits in memory anyway, so Dask is probably not helping you.
it just seems weird to me that
from_array(large_array, chunks=chunks)[0].compute(), for example, does not allocate data to workers the way one would expect from the chunking
in terms of tasks, i'm guessing that the first task (array creation) requires the entire array, so every subsequent task inherits that dependency.
I guess it shouldn't be too hard to kick off dask-jobs from within Flask. Was just wondering whether there are any examples and experiences doing that around.
Not having to setup message queues and the like seems like a huge plus to me.
but thanks for the documentation link
it looks like all these methods target optimal use of data that has been localized to the cluster using e.g.
scatter, is that right?
in my use case, there's no explicit management of data on the cluster -- I just locally define a dask array and try to compute its first element (for example) using the distributed scheduler
Yes, about manully moving things around in cluster memory. https://distributed.dask.org/en/latest/work-stealing.html describes when Dask might do this automatically
You should isolate the exact code, expectations and findings in an issue, then
hello everyone, I have question. I'm trying to run computation that would require large dataset as argument. Something like df.apply(navigate_over_array, args=(very_big_array,)). I'd like to pre-scatter very_big_array to all workers so function can reference it there. What's good way to pass it? If I scatter, I get future, is there a way that would allow me to access data behind this future from applied function in worker?
or is there better way to achieve this result?
I'm configuring Dask on a single node together with a JupyterHub instance. I would like to specify a default LocalCluster when people create one. I found the distributed.yaml in /etc/dask. Can I specify stuff like n_workers, threads_per_worker anywhere there? I've been looking for a complete list of configuration options, but I'm unable to find it
You should be able to find the relevant parts of the graphical output quite easily. It ought to be possible to make aggregate graphics (like the memory and processing sections on the status/ dashboard) or just refuse to work with datasets that have become too large, or even to monitor internal memory usage of the dashboard and halt when it gets too big.
(as I've just done myself :D)
ended up commenting here https://github.com/dask/distributed/issues/2979#issuecomment-530915742
It is, of course, in a separate thread, as opposed to separate process, so it will always be a little tricky to know when some memory limit is approached because of the graphics; however, I can imagine safety safeguards which turn off some graphics known to be resource-hungry above some threshold
Hi all, I'm pretty new to Dask and just had a quick question I was wondering if anyone could help me with - I have two 3D dask arrays and want to broadcast the
solve operation across the last 2 dimensions, i.e. given A and B I want to do C[i,:,:]=solve(A[i,:,:], B[i,:,:]) for all i in the leading dimension... I tried the below but it seems to be really slow (slower than numpy) - does anyone know what I'm doing wrong/what I could do better? Sorry if this was the wrong place to ask.C = da.apply_gufunc(np.linalg.solve, "(i,j),(i,k)->(j,k)", A, B, vectorize=True,output_dtypes=A.dtype)