Fixes langchain-ai/langgraph#8382
`DeltaChannel` rebuilds its value by replaying ancestor writes through
the reducer. Every saver ordered a checkpoint's writes by `(task_id,
idx)`, but live execution applies a super-step's writes in task-path
order (`apply_writes` sorts tasks by `task_path_str(task.path[:3])`).
`task_id` is a hash of the path, so when two or more tasks write one
`DeltaChannel` in the same super-step, replay applies them in an
arbitrary permutation.
Reducers must be batching-invariant, not order-invariant, so the
permutation changes the value: `get_state` disagrees with what `invoke`
returned, and continuing the thread saves the reordered value as the
base for every later write.
## Fix
Replay orders by `(task_path, task_id, idx)`, the order
`SELECT_PENDING_SENDS_SQL` already uses for sends. `apply_writes` and
every saver now sort with one `writes_sort_key` from
`langgraph.checkpoint.base`, in Python, so live and replay can't drift
apart and don't depend on a database collation. `apply_writes` sorts by
the full task path, the string `put_writes` stores, instead of
`path[:3]`. The paths that reach `apply_writes` no longer carry task
ids, so the live order doesn't change.
`BaseCheckpointSaver`'s default `get_delta_channel_history` can't sort
by path, since `PendingWrite` has no task path, so it replays each
checkpoint's writes in `get_tuple` order. `get_tuple` now has to return
`pending_writes` in `writes_sort_key` order, and memory, SQLite and
Postgres do, so the default walk is right for any saver that follows
that. `AsyncPostgresSaver` also gets the sync
`get_delta_channel_history` that `AsyncSqliteSaver` already had; sync
reads used to fall back to the default walk.
Memory and Postgres already stored `task_path`. SQLite accepted it on
`put_writes` and dropped it, so `writes` gains the column and `setup()`
adds it to existing databases.
Writes without a task path (graph input, `update_state` updates, rows
written before the column existed) sort first by `task_id`, which is
also where live execution applies input.
## Compatibility
- Memory and Postgres already had `task_path` on disk, so existing
threads replay in the corrected order after upgrade.
- On SQLite, rows from before the upgrade read back as `''` and keep
their old order. Only new writes get the corrected order, since there is
no stored path to recover.
- SQLite has no `ADD COLUMN IF NOT EXISTS`, so `setup()` runs the
`ALTER` and treats `duplicate column name` as success, which also covers
two processes migrating one file. That error comes while SQLite prepares
the statement, before it asks for the write lock, so `setup()` on an
up-to-date database doesn't wait on other writers. A read-only database
from before the column is read with empty paths instead of failing
setup.
- Downgrade is safe: older code's `INSERT` omits the column and the
column has a default.
- `get_tuple` and `list` return `pending_writes` in `writes_sort_key`
order instead of task id order (SQLite, Postgres) or put order (memory).
Nothing in langgraph depends on the old order.
- Postgres reads `task_path` and `idx` with each pending write, so a
subclass that pairs `SELECT_SQL` with its own `_load_writes`, or the
other way round, has to match the new row shape.
- A third-party saver has to replay in `writes_sort_key` order: in its
own `get_delta_channel_history` if it has one, otherwise by returning
`pending_writes` from `get_tuple` in that order. Otherwise it fails the
new conformance tests; the `get_tuple` one is a base test, so it runs
for every saver.
- langgraph, checkpoint-sqlite and checkpoint-postgres import
`writes_sort_key` from langgraph-checkpoint, so this bumps
langgraph-checkpoint to 4.3.0 and requires it there. 4.3.0 has to be
released first.
## Limits
- Several updates in one `bulk_update_state` super-step share a task id,
so the saver keeps only the first one's writes and the `DeltaChannel`
loses the rest (plain channels are fine). That's on main too; #9128
fixes it and stores a task path per update, so with this PR those
updates also replay in the order given.
- Exit durability stores its writes differently; resume ordering there
is fixed in #9114, stacked on this PR.
- langgraphjs applies `DeltaChannel` writes in task id order, live and
on replay (langchain-ai/langgraphjs#2544), so its reads match too.
Moving JS to task path order is a follow-up.
## Tests
Each fails on main and passes here:
- `test_delta_channel_parallel_order.py`, across memory, SQLite and
Postgres: parallel writers and a `Send` fan-out, comparing `get_state`
with the live result, including sync `get_state` on the async savers.
- Three conformance tests whose task ids sort opposite to their task
paths, so they cannot pass by chance: two on
`get_delta_channel_history`, one on `get_tuple`'s `pending_writes`
order.
- SQLite migration tests from a pre-column database, sync and async,
including a read-only file.
One more that passes on main too: SQLite `setup()` while another
connection holds the write lock, with no busy timeout, sync and async.
It keeps the `ALTER` from waiting on that lock.
Thanks to @ErenAta16 for the report, the reproduction, and for tracing
`task_path` through all three backends, and to @ragnarok268 for the
conformance-test design with task ids ordered opposite to their paths.
---------
Co-authored-by: ErenAta16 <149434812+ErenAta16@users.noreply.github.com>
Co-authored-by: ragnarok268 <58264829+ragnarok268@users.noreply.github.com>
