Issue with Checkpoint Persistence in LangGraph Docker Deployment

[BoyPit]

Hi,

I'm currently building a LangGraph container using the CLI. Everything works well alongside Redis and PostgreSQL containers.

However, after redeploying the stack, I sometimes lose my checkpoints — even though I haven’t touched the PostgreSQL volume.

I'm unsure whether I'm managing checkpointing correctly with LangGraph Server in Docker. Currently, I'm using:

checkpointer = MemorySaver()
graph = workflow.compile(checkpointer=checkpointer)

Does this store checkpoints in the PostgreSQL container at all?

Should I instead just use:

graph = workflow.compile()

Or would it be better to explicitly use a PostgresSaver, as described in the https://api.python.langchain.com/en/latest/checkpoint/langchain_postgres.checkpoint.PostgresSaver.html

Thanks for your help!

[hinthornw]

Hi @BoyPit, thanks for writing in. The LangGraph API actually ignores any custom checkpointer when you’re running on LangGraph Server (Docker or Cloud). I'd recommend dropping the checkpointer=MemorySaver() argument and just using the built-in postgres checkpointer that ships with the platform (by compiling with no args graph = workflow.compile())

If you want to point to another custom Postgres instance instead of the default, simply set the POSTGRES_URI environment variable before you start the server:

export POSTGRES_URI="postgresql://user:pass@postgres:5432/mydb"

More details here: https://langchain-ai.github.io/langgraph/cloud/reference/env_var/?h=environ#postgres_uri_custom

With that in place, LangGraph Server will persist all checkpoints into your Postgres volume and they’ll survive container restarts.

Note that "MemorySaver" is an "In memory" implementation of the checkpointer interface, meaning that if we were to actually use your code as-is, nothing would actually be persisted. This is why we frequently sayh to use the in-memory saver only for testing & debugging purposes example: here

[danni-boii]

I have experienced the same loss on langgraph checkpointers at my local dev environment.

As to my memories, earlier versions of InMemorySaver can actually persist the checkpointers data even if it is under local dev in a folder named .langgraph_api by default.

I believed it's a bug that the current InMemorySaver is having issues reading checkpointers data back from the local persisted pickle files.

[BoyPit]

I have experienced the same loss on langgraph checkpointers at my local dev environment.

As to my memories, earlier versions of InMemorySaver can actually persist the checkpointers data even if it is under local dev in a folder named .langgraph_api by default.

I believed it's a bug that the current InMemorySaver is having issues reading checkpointers data back from the local persisted pickle files.

Thanks !
I checked and it's a bug for Pg. You need to upgrade langgraph-checkpoint if you plan to use pg checkpoint with custom URI.

[ivrschool]

@hinthornw
Thanks for sharing this. I have checked it and it does not work.
export POSTGRES_URI="postgresql://user:pass@postgres:5432/mydb"

Is there any alternative to this?

[iliox]

I am using pants and when I am trying to deserialise the checkpointer, I am losing channel_values from the checkpointer state.
This works fine in local python 3.11 env but doesn't work with pants for some reason. This is my serialization logic:

def serialize_checkpointer(checkpointer: MemorySaver, config) -> str:
    checkpointer_state = checkpointer.get_tuple(config)
    return base64.b64encode(pickle.dumps(checkpointer_state)).decode("utf-8")


def deserialize_checkpointer(serialized_state: str = "") -> MemorySaver:
    if serialized_state == "":
        return MemorySaver()
    tuple = pickle.loads(base64.b64decode(serialized_state.encode("utf-8")))
    checkpointer = MemorySaver()
    checkpointer.put(
        config=tuple.parent_config,
        checkpoint=tuple.checkpoint,
        metadata=tuple.metadata,
        new_versions={},
    )
    return checkpointer