langgraph-checkpoint-dmdb 0.1.1

DM Database checkpoint and Store persistence for LangGraph.

LangGraph Checkpoint DMDB

LangGraph checkpoint and Store persistence for DM Database, using dmPython.

Compatibility

• langgraph-checkpoint>=2.1.1,<2.2
• dmpython==2.5.32
• Python 3.9+

The package provides:

• DMDBSaver: complete checkpoint history and time travel.
• ShallowDMDBSaver: latest-checkpoint-only persistence.
• AsyncDMDBSaver and AsyncShallowDMDBSaver: async graph APIs backed by isolated worker-thread connections.
• DMDBStore and AsyncDMDBStore: LangGraph Store CRUD, batch, JSON filtering, pagination, and namespace listing.

Installation

pip install langgraph-checkpoint-dmdb

Usage

For a long-lived compiled graph:

from langgraph.checkpoint.dmdb import DMDBSaver

saver = DMDBSaver.from_conn_string(
    "dmdb://sysdba:password@localhost:5236/NLP_KNOWLEDGEBASE"
)
saver.setup()
graph = builder.compile(checkpointer=saver)

config = {"configurable": {"thread_id": "conversation-1"}}
for event in graph.stream(inputs, config=config):
    ...

Applications that already have structured database configuration should avoid constructing a URI:

saver = DMDBSaver.from_conn_params(
    host=db.host,
    port=int(db.port),
    user=db.username,
    password=db.password,
    schema=db.database,
)
saver.setup()
graph = builder.compile(checkpointer=saver)

Short-lived scripts may use the Saver as a context manager:

with DMDBSaver.from_conn_string(DMDB_URI) as saver:
    saver.setup()
    graph = builder.compile(checkpointer=saver)
    result = graph.invoke(inputs, config=config)

from_conn_string() and from_conn_params() create a connection factory. Each checkpoint operation owns a new dmPython connection and closes it afterward, so a Saver can be shared by concurrent synchronous graph executions without sharing a connection.

Do not return a graph from inside the with block and use it after the block exits; the Saver is closed on context exit.

Async Checkpointing

dmPython has no native async API. The async savers run each database operation with asyncio.to_thread() and obtain a fresh connection from the factory, so connections are never shared between worker threads:

from langgraph.checkpoint.dmdb import AsyncDMDBSaver

async with AsyncDMDBSaver.from_conn_params(
    host=db.host,
    port=int(db.port),
    user=db.username,
    password=db.password,
    schema=db.database,
) as saver:
    await saver.setup()
    graph = builder.compile(checkpointer=saver)
    result = await graph.ainvoke(inputs, config=config)

Async savers require a connection factory. Passing one shared dmPython connection is rejected because dmPython.threadsafety == 1.

Shallow Saver

from langgraph.checkpoint.dmdb import ShallowDMDBSaver

saver = ShallowDMDBSaver.from_conn_string(DMDB_URI)
saver.setup()
graph = builder.compile(checkpointer=saver)

The shallow saver retains only the latest checkpoint per thread and namespace. Do not use shallow and full savers for the same thread/namespace because shallow writes intentionally prune that history.

LangGraph Store

from langgraph.store.dmdb import DMDBStore

store = DMDBStore.from_conn_params(
    host=db.host,
    port=int(db.port),
    user=db.username,
    password=db.password,
    schema=db.database,
)
store.setup()
store.put(("users", user_id), "preferences", {"language": "zh-CN"})
items = store.search(("users", user_id), filter={"language": "zh-CN"})
graph = builder.compile(checkpointer=saver, store=store)

AsyncDMDBStore exposes the corresponding aget, aput, adelete, asearch, alist_namespaces, and abatch APIs.

Connection URI

Supported schemes are dmdb:// and dm://:

dmdb://user:password@host:5236/schema

The path is treated as the DM schema. The Saver validates it and executes SET SCHEMA after connecting. Usernames and passwords containing URI-special characters must be percent encoded. from_conn_params() does not require URI encoding.

If dmPython raises DatabaseError: [CODE:-70089] Encryption module failed to load in Linux, add the wheel's bundled dmssl directory to LD_LIBRARY_PATH before starting Python. The exact site-packages path depends on the runtime image; see the dmPython 2.5.32 package metadata for its supported SSL setup.

Transactions and concurrency

• dmPython connections use autoCommit=False.
• put, put_writes, and delete_thread commit atomically and roll back on error.
• Factory-backed Savers create an independent connection per operation.
• A directly supplied dmPython connection remains owned by the caller and is restricted to its creating thread.
• setup() is idempotent and must be called before first use.
• DM MERGE source parameters are explicitly cast to their target SQL types. BLOB/write batches execute row-by-row inside one transaction because dmPython/DM otherwise cache the first inferred bind length and can raise -70005 String truncated for later values.

Driver Limitations

• Async APIs use worker threads rather than a native async DM protocol.
• Async classes cannot accept a single direct dmPython connection.
• Store JSON filtering and namespace matching run in Python to avoid DM-version-specific JSON functions.
• Store semantic query and vector indexing are not implemented; the MySQL 2.0.17 Store also did not implement semantic indexing.

Tests

Unit and LangGraph integration tests do not require a DM instance:

uv run pytest -m "not integration"

Real DM integration requires dmPython and these environment variables:

DM_HOST
DM_PORT
DM_USER
DM_PASSWORD
DM_SCHEMA

Run it with:

uv run pytest -m integration

Design

The implementation plan and release gates are documented in docs/dmdb-compatibility-plan.md. The complete MySQL-to-DMDB feature mapping is in docs/mysql-feature-parity.md.

This project derives serialization behavior and tests from langgraph-checkpoint-mysql 2.0.17. The original MIT copyright notice is retained in LICENSE.