The project is still crating many gcs clients

[david-gang]

After i fixed: ef44612

I suddenly saw that ImageBytesLoader is created ad hoc in every function. So the commit mentioned is just a first step to fix this issue.

Look please here: https://github.com/search?q=repo%3Alangchain-ai%2Flangchain-google%20ImageBytesLoader&type=code

What do you suggest here?

The path which i hit is at

langchain-google/libs/vertexai/langchain_google_vertexai/chat_models.py

Line 221 in 7f704aa

return ImageBytesLoader(project=project).load_gapic_part(path)

and this is a stateless function.

I see one possible solution which is creating a cache of imagebytsloader. for xample creating a map from project to imagbytesloader (lru_cache) . But this is a bigger change and want to hear what you think.

[lkuligin]

maybe it's easier to create a property and a lazily initialzed _image_loader attribute (the first time you get the property, it checks whether the loader is initialized, and initializes it if needed)

[jzaldi]

Why is this a problem? AFAIK GCS Client is just a bundle of configuration parameters to make API calls (but feel free to correct me). Is this a big performance hit?

[david-gang]

as far as i know every storage client maintains its http client which is a request session so it will not reuse connections

[david-gang]

maybe it's easier to create a property and a lazily initialzed _image_loader attribute (the first time you get the property, it checks whether the loader is initialized, and initializes it if needed)

I don't understand what this means exactly. I have two issues here:

• The class is created from a function and not another class so i don't understand the attribute
• We are passing project as input so an attribute is not enough we need a map from project to client

[jzaldi]

Did a dirty performance check and it seems that it does have an impact.

I think what @lkuligin means is:

• Make storage_client a ImageBytesLoader attribute (Maybe optional?)
• Have a client (lazyly initialized and cached) at ChatVertexAI level
• Pass this client in the builder of ImageBytesLoader so that it reuses it.

This limits the client creation to one per ChatVertexAI initialization