Files
langchain/libs/core/tests
WU LIFU ad16eed119 core[patch]: runnable config ensure_config deep copy from var_child_runnable… (#24862)
**issue**: #24660 
RunnableWithMessageHistory.stream result in error because the
[evaluation](https://github.com/langchain-ai/langchain/blob/master/libs/core/langchain_core/runnables/branch.py#L220)
of the branch
[condition](99eb31ec41/libs/core/langchain_core/runnables/history.py (L328C1-L329C1))
unexpectedly trigger the
"[on_end](99eb31ec41/libs/core/langchain_core/runnables/history.py (L332))"
(exit_history) callback of the default branch


**descriptions**
After a lot of investigation I'm convinced that the root cause is that
1. during the execution of the runnable, the
[var_child_runnable_config](99eb31ec41/libs/core/langchain_core/runnables/config.py (L122))
is shared between the branch
[condition](99eb31ec41/libs/core/langchain_core/runnables/history.py (L328C1-L329C1))
runnable and the [default branch
runnable](99eb31ec41/libs/core/langchain_core/runnables/history.py (L332))
within the same context
2. when the default branch runnable runs, it gets the
[var_child_runnable_config](99eb31ec41/libs/core/langchain_core/runnables/config.py (L163))
and may unintentionally [add more handlers
](99eb31ec41/libs/core/langchain_core/runnables/config.py (L325))to
the callback manager of this config
3. when it is again the turn for the
[condition](99eb31ec41/libs/core/langchain_core/runnables/history.py (L328C1-L329C1))
to run, it gets the `var_child_runnable_config` whose callback manager
has the handlers added by the default branch. When it runs that handler
(`exit_history`) it leads to the error
   
with the assumption that, the `ensure_config` function actually does
want to create a immutable copy from `var_child_runnable_config` because
it starts with an [`empty` variable
](99eb31ec41/libs/core/langchain_core/runnables/config.py (L156)),
i go ahead to do a deepcopy to ensure that future modification to the
returned value won't affect the `var_child_runnable_config` variable
   
   Having said that I actually 
1. don't know if this is a proper fix
2. don't know whether it will lead to other unintended consequence 
3. don't know why only "stream" runs into this issue while "invoke" runs
without problem

so @nfcampos @hwchase17 please help review, thanks!

---------

Co-authored-by: Lifu Wu <lifu@nextbillion.ai>
Co-authored-by: Nuno Campos <nuno@langchain.dev>
Co-authored-by: Bagatur <baskaryan@gmail.com>
2024-08-01 17:30:32 -07:00
..