Open xKHUNx opened 4 years ago
I would like to request for the feature to use the custom serializer, as set from the Client object.
I agree that that would be useful in many cases. One would need to think about how to communicate that custom serializer to all of the parts of the cluster. There is also an issue in that many people use the serializers functionality in order to restrict custom serialization. For example some people don't want to allow pickle deserialization in the scheduler for security reasons.
I think that in the end security will probably take precedence over convenience, and you'll be asked to specify your serializer on each Worker/Scheduler when you set them up.
I'm having a use case which requires me to parallelize nested Cython functions using Dask, and from the documentation, I've learned that the we can set the serializers/deserializers for a Client object. This means I can implement my own serializer, in this case, dill, to parallelize my nested functions. It looks like this.
However, when I'm using
map()
from the client, I am still unable to serialize it, and upon inspecting the error message, it turns out it is using pickle (which will not work for nested function) indumps_function()
to serialize my function in the class.Here's the error message:
I would like to request for the feature to use the custom serializer, as set from the Client object.