Hi team,
How can I conduct parallel IO with kedro...
# questionsb
Baden Ashford
07/31/2023, 11:16 AMHi team,
How can I conduct parallel IO with kedro? I have a larger than memory partitioned dataset. I'd like to run each partition through the node in some parallel fashion. Can I utilise ParallelRunner for this?
Thank you 😁
Baden Ashford
07/31/2023, 3:59 PMI have found this post regarding a custom dask runner https://docs.kedro.org/en/stable/deployment/dask.html but it does not seem like a straightforward thing to drop into an otherwise series pipeline
Baden Ashford
08/01/2023, 9:40 AMSorry to bug you, but any ideas @datajoely @Nok Lam Chan?
Chat GPT recommends to use dask within the node function, but idk how this sits with the 'nodes as pure functions' paradigm.
Copy code
def process_single(df: pd.DataFrame) -> pd.DataFrame:
df = process_thing(df=df)
return df
def process_batch(parts: dd.DataFrame) -> dd.DataFrame: # this is the node
dask_client = Client()
batch = parts.map_partitions(process_single, meta=parts._meta)
return batchd
datajoely
08/01/2023, 9:45 AMyeah that’s not the way we’d do it
😁 1
datajoely
08/01/2023, 9:46 AMare the instructions on the linked docs not clear?
datajoely
08/01/2023, 9:47 AMessentially we tend to push the remote execution engine connection point to the DataSets, Hooks or Runner. The nodes shouldn’t be aware of IO
b
Baden Ashford
08/01/2023, 9:48 AMThe parallel runner docs or the dask runner docs?
d
datajoely
08/01/2023, 9:49 AMdask runner
datajoely
08/01/2023, 9:49 AMI would say on the regular parallel runner have you tried returning generators?
datajoely
08/01/2023, 9:50 AMdatajoely
08/01/2023, 9:50 AMlazy saving may be the answer here
b
Baden Ashford
08/01/2023, 9:55 AMOk thanks for the link I will take a look 🙏
To answer your questions, the dask runner docs isn't super clear on how this runner would allow me to map over partitions of a dask dataframe and process them in parallel but idk if this is asking too much of it. And I have no returning generators but I have seen it mentioned in other threads so can take a look as well and report back 👍
d
datajoely
08/01/2023, 9:58 AMthe returning generators wont run in parallel but they hopefully will allow you to process larger than memory partitions
datajoely
08/01/2023, 9:58 AMif you use Dask or Spark you’re delegating to those engines for task splitting and scale out
datajoely
08/01/2023, 9:59 AMthe last idea is to use Polars which should be even simpler
n
Nok Lam Chan
08/01/2023, 10:35 AMSome answered thread before: https://kedro-org.slack.com/archives/C03RKP2LW64/p1689857227419389?thread_ts=1689789306.348409&cid=C03RKP2LW64
Nok Lam Chan
08/01/2023, 10:35 AMThere isn’t a single solution fit all cases, thus you see a tons of dataframe libraries, pandas, polars, dask, spark etc. It’s a spectrum of problem.
Nok Lam Chan
08/01/2023, 10:37 AMI have a larger than memory partitioned dataset.It also depends on how much larger it is, can you solve it by processing it in chunks? In general, I’ll only bring in distributed system when you are running out of solution with a single machine.
Nok Lam Chan
08/01/2023, 10:38 AMlazy saving may be the answer hereI agree with this, I am working with an example here but it will take some time
b
Baden Ashford
08/01/2023, 12:04 PMNice thank you for the links, I think I am prematurely optimising, thanks for your help!
K 1
2 Views