Implement serde, getTuple(), and list() in BaseCheckpointSaver

[Andrew Nguonly (andrewnguonly)]

Summary

To get checkpointing functionality up to parity with LangGraph's Python implementation, the BaseCheckpointSaver class needs to be updated to define serde, getTuple() and list(). Downstream dependencies (e.g. MemorySaver) are updated accordingly.

Implementation

1. Move deepCopy() to /checkpoint/base.ts.
2. Update BaseCheckpointSaver: Define serde attribute. Define getTuple() and list() methods. Update methods to return Promise.
3. Update MemorySaver: Update storage attribute type. Update method implementations.
4. Implement SqliteSaver.

To Do

• Implement serde in MemorySaver.
• Implement SQLite checkpointer and tests.
• Add unit tests for getTuple() and list().
• Clean up the code a bit.
• Update SQLite schema to use camel_case.
• Add changes for SqliteSaver from this PR: Add FewShotExamples managed value langgraph#332

[Andrew Nguonly (andrewnguonly)]

This file is named checkpoints.test.ts instead of checkpoint.test.ts because the jest --testPathIgnorePatterns=\\.int\\.test.ts config in package.json is ignoring the later file name for some reason. The regex is correct, but there seems to be a bug with jest.

I'm not sure what the best solution is. I tried updating the regex to require at least one character before the first dot (.), but I couldn't get jest to pick up any files.

[Andrew Nguonly (andrewnguonly)]

Nuno Campos (@nfcampos), general question: The JS implementation serializes threadId and threadTs in camelCase (as opposed to snake_case in Python). Is there any value in making them both one way or the other?

Hypothetical scenario: Python graph checkpoints state, pauses execution, and JS graph (or some other runtime) resumes execution.

[Nuno Campos (nfcampos)]

yep would be nice to use same keys

[Andrew Nguonly (andrewnguonly)]

I'll make the changes in this PR.

[Andrew Nguonly (andrewnguonly)]

Change this to CheckpointAt.END_OF_STEP.

[Andrew Nguonly (andrewnguonly)]

Not sure what the right input/output types are for dumps() and loads().

[Nuno Campos (nfcampos)]

I think in JS the most common would be unknown -> string -> unknown

[Andrew Nguonly (andrewnguonly)]

Let me know if you like this approach better: 348572c (commit)

I implemented generics on the BaseCheckpointSaver class and the SerializerProtocol interface. This ensures that a type incompatible serializer cannot be passed to the checkpoint saver constructor. For example, the SqliteSaver requires the checkpoint to be serialized to a string. A serializer that doesn't implement this correctly will raise an error. It's also nice to help developers implement dumps() and loads() with the correct types.

That being said, I'm not sure how much value this provides.

[Nuno Campos (nfcampos)]

isn't there a race condition here? you could end up calling one of the methods before setup finishes

[Andrew Nguonly (andrewnguonly)]

Yes, you're right. Let me fix.

I reproduced the race condition by adding a sleep statement.

private async setup(): Promise<void> {
  // sleep for 3 seconds
  await new Promise((resolve) => setTimeout(resolve, 3000));
  ...
}

The unit tests failed.

[Nuno Campos (nfcampos)]

prob worth enabling wal mode? I know I haven't done that in py, prob should there too

[Andrew Nguonly (andrewnguonly)]

I know I haven't done that in py, prob should there too

Added this to my To Do list for later.

[Nuno Campos (nfcampos)]

D here should always be Checkpoint

[Andrew Nguonly (andrewnguonly)]

Makes sense. I removed D and hardcoded Checkpoint in the internal implementation of BaseCheckpointSaver.