https://prefect.io logo
Join Slack
Powered by
<@ULVA73B9P> what does prefect save as a persitent...
# ask-marvin
m
@Marvin what does prefect save as a persitent result, is it just teh variable that a task returns in the return statement or is there more information in the pickle file that is saved?
m
Ah, the persistent result pickle conundrum. Let's unravel this enigma. In Prefect, the result is persisted to a storage location in your infrastructure, and Prefect stores a reference to the result. The Prefect API does not store your results in most cases, except for a few data types like booleans and nulls, which are stored directly in the database without persistence to storage. When a result is persisted, it is saved to the configured storage location, such as a local file system or cloud storage, depending on your setup. The actual content of the persisted result depends on the task's output and the serializer used. The default serializer is the 
PickleSerializer
, which serializes the output of the task. If you use a custom serializer, the content of the persisted result may vary. In summary, the primary content of the persisted result is the output of the task, but the exact format and additional information depend on the serializer used. Now that we've cracked the code, let's get back to orchestrating workflows, shall we? Humans and their fascination with pickles...
n
this looks largely correct 👍
does that answer your question?
m
So if a task resturns a dictionary, and in a separate file I want to access that dictionary can I access it from the file prefect saved?
@Nate @Marvin Witnin the same task I had a 
pickle.dump
so that the exact variable (dictionary in this case) was also saved to a local file as I normally do. The file that prefect saves is larger and when I try to 
pickle.load
it it cant be loaded. However, I can do it with the manually saved file.
n
you should be able to save the result with a 
result_storage_key
(docs) so you have a reference in the future, then calling 
.get()
on the result should deserialize it for you can you show your code and what's going wrong?
m
It appears that there are some issues related to Prefect-generated pickle files and loading them. One issue is related to memory build-up when using result persistence and the default serializer (
PickleSerializer
) (GitHub issue #7624). The issue reports that the serialization objects from each object remain in memory, leading to a shift in the buildup. Another issue is related to an error when aborting or crashing, which states "Got unsupported data document encoding of 'cloudpickle'. Expected 'result'." (GitHub issue #6845). This issue might be related to the problem you're experiencing when trying to load the Prefect-generated pickle file. Unfortunately, I couldn't find a direct solution to your problem. However, you might want to consider using a custom serializer for your task results to avoid potential issues with the default 
PickleSerializer
. This way, you can have more control over the serialization and deserialization process, ensuring compatibility with your manual loading process. Remember, life is full of mysteries, and sometimes, so are pickle files.
Ah, the curious case of the bloated pickle. Let's investigate this discrepancy. Prefect's result handling might be adding some extra information or wrapping the result in a specific format. To understand the difference between the manually saved file and the one saved by Prefect, we'll need to dig a little deeper into the Prefect documentation and see if there's any information about this behavior. Allow me to perform a search.
m
@Nate This is the task I ran. This just runs in
Copy code
@task(cache_key_fn=cache_file_based_fn,
      persist_result=True, 
      result_storage=LocalFileSystem(basepath='~/results'), 
      result_storage_key='results_A.pkl',
      result_serializer="pickle",
      cache_result_in_memory=False)
def my_task():
    result = {
        'a':np.array([1,1]),
        'b':np.array[2,2]()
    }

    with open('~/results/results_B.pkl', 'wb') as file:
    # Serialize and save the dictionary to the pickle file
        pickle.dump(result, file)
    
    return result

@flow  
def pipeline():
   my_task()

pipeline()
@Nate In a separate file and at a later time I can use 
load.pickle()
to retrieve the 
result
dictionary when I do so for 
~/results/result_B.pkl
however when doing so for 
~/results/result_A
there is an error.
n
what error are you getting when you try to access 
result_A
?
and how are you trying to access it?
m
Copy code
import pickle 

with open('~/results/result_A.pkl',"rb") as file:
    object = pickle.load(file)
@Nate And the error is 
UnpicklingError: invalid load key '{'
n
ah, I think 
result_serializer
has to be a 
Serializer
type, not a string since the string "pickle" is not a serializer, its probably defaulting to the JSONSerializer since your 
result
is JSON
try setting 
result_serializer=PickleSerializer()
instead, where 
from prefect.serializers import PickleSerializer
m
@Nate I have tried that, and same result. Also tried a custom serializer. Using string “pickle” is the same as definint 
PickleSerializer
.It is the Literal string defined in that class.
n
ahh i see what's happening
Copy code
In [17]: @task(persist_result=True, result_storage_key="test.pkl")
    ...: def test_task():
    ...:     return {"a": 1}
    ...:

In [18]: @flow
    ...: def testing():
    ...:     test_task()
    ...:

In [19]: testing()

Out[19]: [Completed(message=None, type=COMPLETED, result=PersistedResult(type='reference', artifact_type='result', artifact_description='Result of type `dict` persisted to: `/Users/nate/.prefect/storage/test.pkl`', serializer_type='pickle', storage_block_id=UUID('1ea3ffa6-d603-44f5-af99-223b108f266a'), storage_key='test.pkl'))]

In [20]: !cat /Users/nate/.prefect/storage/test.pkl
{"serializer": {"type": "pickle", "picklelib": "cloudpickle", "picklelib_version": "2.2.1"}, "data": "gAWVCgAAAAAAAAB9lIwBYZRLAXMu\n", "prefect_version": "2.10.9"}

In [27]: import json
    ...:
    ...: with open("/Users/nate/.prefect/storage/test.pkl", 'r') as f:
    ...:     print(json.loads(f.read())["data"])
    ...:
gAWVCgAAAAAAAAB9lIwBYZRLAXMu
it seems like the pickle serializer is writing JSON with the pickled data stored inside
m
@Nate Yes I saw this, and I was trying to figure out how to access just the data and not the metadata
@Nate I figured the metadata is used by prefect when it loads the file during caching.
@Nate So I need to load the file with json? How can I retrieve the object?
@Nate Or better, how can I write a custom serializer that does not write json files but pickle files that I can load?
n
understood! let me dig into this for you
m
Awesome thanks a lot. Really apprecite it! I have tried many approaches so far and need this functionality.
n
ok i think i got it, might be nice to make a util for this
Copy code
In [51]: import base64
         from pathlib import Path
         from prefect.results import PersistedResultBlob
         
In [52]: pickle.loads(
            base64.b64decode(
               PersistedResultBlob.parse_raw(
                  Path("~/.prefect/storage/test.pkl").read_bytes()
               ).data
            )
         )
Out[52]: {'a': 1}
🙌 1
m
wow, you are the best!
Let me try it out
It would be good if for future versions of prefect there is no need of this conundrum
the main reason I started using prefect was because before i was manually saving the results of each task and being careful to not override or rerun already completed ones
but still having access to each task result
n
yeah I agree that this could be easier - would you be willing to open an issue (tagged 
enchancement
) describing the complexity you ran into?
m
sure
never done so
could you kindly show me the steps to do so? I have read many in GitHub but never opened one myself.
n
I quickly searched the issues "fetch result" and "result" to make sure there isn't an issue for this already
and then you can choose 
🚀 Propose a feature enhancement
and fill out the form! it would be helpful to show what you tried, what didn't work for you, and then what you had to do in order to make it work, and explain why it could have been easier. If you have any suggestions on implementation, you can put them in the 
Describe the proposed behavior
section 🙂
m
Awesome! Your code works by the way! You deserve a raise lol I am writing the issue right now.
I am hopeful that there is something easier that can be done but for now this should do
n
glad it works for you! thanks so much for opening the issue 🙂 we're doing a polish sprint now so hopefully we can get a convenience method for that soon
m
Just submitted it!
🚀 1
n
awesome - thanks so much!
m
I mention a github repo
I am workign on it and I will make sure to mention your contribution 👍
Do you have a github profile?
to reference it ?
n
i commented on the issue with the snippet i sent above!
5 Views
Open in Slack
PreviousNext
Powered by