Can I skip checkpointing for subgraphs?

[brayMaersk]

Hi everyone 👋

I'm exploring ways to improve the overall experience of our agentic solution and one idea I’m considering is skipping the checkpointer for our RAG subgraph.

Because of the current solution the checkpointer in Mongo is growing. I am thinking of excluding the RAG search output from the checkpointer by removing the checkpointing from the RAG subgraph as the subgraph is primarily responsible for retrieval and doesn’t carry critical state between steps, avoiding checkpointing here might reduce the checkpointer size without impacting reliability.

Before I proceed, I wanted to check:

Has anyone tried this approach in production?
Are there any hidden pitfalls or best practices I should be aware of?
Would this impact observability or rollback in any significant way?
Appreciate any thoughts or experiences you can share. Thanks in advance!

[onestardao]

skipping checkpoints in a rag subgraph can work, but you’re basically stepping into what i’d tag as a state continuity blindspot — in my own problem index it’s close to No.7.
it’s true the retrieval step doesn’t “own” the critical state, but once you skip logging here, you lose the ability to detect when upstream context drifted or when retrieval silently failed but still returned “something”. in production, those silent state drops are nasty because rollback only looks clean until you try to reproduce the original execution.
i’ve got a checklist that covers this exact trade-off (size vs reliability) and how to keep observability without bloating the mongo checkpoint store. if you want it, happy to share.