Feasibility of Using In-Memory Checkpointer for Prod environment

[Subham07]

I want to use the "interrupt_before" logic for langgraph agent, for an agent in production environment where the user input will decide whether to execute a tool or not.

For the same purpose I am using checkpointer and thread_ids to get the state and resume the graph incase the approval is made.
I am also following this notebook : https://github.com/langchain-ai/langgraph/blob/main/docs/docs/how-tos/human_in_the_loop/breakpoints.ipynb

Now I wanted to ask if keep using the checkpointers, will it cause my application to break because of memory usage as it will keep storing the snapshots in a dictionary?

Also in this documentation: https://langchain-ai.github.io/langgraph/reference/checkpoints/#checkpointer-implementations
its written that "Since checkpoints are saved in memory, they will be lost when the program exits. Only use this saver for debugging or testing purposes."

Is there any other way to destroy the threads so that memory gets cleared incase there is a problem in future?

[hinthornw]

Checkpointers backed by a production-grade database (such as postgres, redis, etc.) are designed for prod and used at scale today.

That line is in the section on the "MemorySaver", a checkpointer implementation which is in-memory used primarily for proving out solutions locally, using in tests, and use in the docs (since it requires no additional setup). It offers no real "persistence", since it is just a python dictionary.

[Subham07]

thanks for the reply @hinthornw . So if I just use MemorySaver as checkpoint , will it pose any problem in future as using postgres/redis will require extra effort? MemorySaver itself is solving the use-case of interrupt based graph agent.