Closed Bowen0729 closed 2 years ago
Hi @Bowen0729, that is good news to hear. We always want DataPrep to have more ecosystem integrations. May I ask if you can write down how you run DataPrep on the Yarn cluster and then we can convert that into a page in our docmentation?
Sure!@dovahcrow
Install dask-yarn with pip
pip install dask-yarn
Ensure that the libraries used on the Yarn cluster are the same as what you are using locally.
using conda-pack package a conda environment
conda-pack
and upload the archive to HDFS
hdfs:///mypath/archive.tar.gz
Run dataprep eda with the following:
from dask_yarn import YarnCluster
from dataprep.eda import create_report
from dask.distributed import Client
from dask.dataframe as dd
cluster = YarnCluster(environment='archive.tar.gz', worker_memory='10GiB', worker_vcores=4, scheduler_memory='1GiB')
cluster.scale(4)
client = Client(cluster)
ddf = dd.read_parquet(hdfs:///data-path/data.parquet)
create_report(ddf)
Spark supports multiple cluster manager, such as standalone, mesos, hadoop yarn or kubernetes, and I think dataprep based on dask can handle bigdata, which is the advantage over other frameworks, so does dataprep eda need to support other cluster manager? and it will be more friendly to bigdata scenarios, what do you think?
If it is necessary, we can talk about how to design the dataprep on yarn, perhaps user can choose the running mode.
If it is not necessary, I will open a pr for dataprep on yarn docmentation after you verified, and it's my pleasure to be a contributor of dataprep
Hi @Bowen0729 , thanks a lot for the detailed steps! Currently we do not have enough people to make dataprep work on Yarn, which needs many optimizations and testings.
It would be very nice if you could add the doc for Yarn! You could add a section about Yarn in this file: https://github.com/sfu-db/dataprep/blob/develop/docs/source/installation.rst and then open a PR. Thanks for being a contributor of dataprep!
Thank you for reply @jinglinpeng I will finish it in next few days!
My company use Hadoop eco system for bigdata, means that we have a Yarn cluster without a dask cluster. As I know, dask can run on Yarn,recently,I tried to run dataprep on yarn, and it worked well. So, does the community consider support dataprep on yarn? and we can work on it toghter.