-
Notifications
You must be signed in to change notification settings - Fork 55
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
WIP: Expose some information about notebook execution state #79
base: main
Are you sure you want to change the base?
Conversation
Great, I think it would be interesting to have. We could also attach timing information to these events? |
What kind of timing information? Like when a particular state transition happened? |
Yes, we already have some timing information in the cell metadata, so this would cover the rest of the execution process. |
@davidbrochart I've made an update. You mean something like that? |
Yes, and I guess we don't need |
good point; removed. |
nbclient/client.py
Outdated
@@ -32,6 +33,15 @@ def timestamp(): | |||
return datetime.datetime.utcnow().isoformat() + 'Z' | |||
|
|||
|
|||
class ExecutionState(enum.Enum): | |||
NOTHING = 0 | |||
STARTUP = 1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Maybe IDLE
instead of NOTHING
, STARTING_UP
instead of STARTUP
?
nbclient/client.py
Outdated
@@ -323,7 +323,7 @@ def reset_execution_trackers(self): | |||
self.output_hook_stack = collections.defaultdict(list) | |||
# our front-end mimicing Output widgets | |||
self.comm_objects = {} | |||
self.state_history = [] | |||
self.state_history = [ExecutionState.IDLE, timestamp()] |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It must be self.state_history = [(ExecutionState.IDLE, timestamp())]
. Maybe we should add a basic test?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
hah I noticed this immediately after pushing. I will add some state checking to tests regardless
nbclient/client.py
Outdated
@@ -512,7 +534,7 @@ async def async_execute(self, reset_kc=False, **kwargs): | |||
info_msg = await self.async_wait_for_reply(msg_id) | |||
self.nb.metadata['language_info'] = info_msg['content']['language_info'] | |||
self.set_widgets_metadata() | |||
|
|||
self._update_state(ExecutionState.COMPLETE) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Should we add self._update_state(ExecutionState.IDLE)
just after that?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I guess idle and complete have some overlap, but they're still separate useful states, right?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, IDLE
always follows COMPLETE
, except for the first state.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
makes sense.
So I'm a little late to the conversation here, but one thing to consider is having the hook and responding to status more in papermill where there's more of a pattern of user registered control and plugin capability on top of nbclient. I don't see any issue with tracking current state here as done with the enum pattern though. Would need some tests here for a merge. |
Hooks IMO is the superior solution here. I am curious though, why the plugin-friendliness belongs to papermill rather than some mixin or something in nbclient. Is it intended for people to extend nbclient with inheritance? If so, are the public (by convention, things not starting with an underscore) members stable enough to depend on? Or should I wait till 1.0.0? @MSeal I see that you also maintain or at least contribute to papermill, so that's why it's possible there. Generally, if I subclassed nbclient and extended things to put hooks in places I need them, would that be a good long-term solution? If so, I'd close this issue. #81 also referencing this issue which is asking for similar stuff. |
Generally yeah. Or by calling the methods from higher order logical constructs that have further aims. e.g. https://github.com/nteract/testbook (alpha) is using nbclient but to support all sorts of other execution pattern requirements by organizing it's actions in wrapper classes. The difference is that nbclient is the low level primitive library with the encapsulation of cell execution captured. Papermill is an opinionated library with plugin registry systems and flexibility for higher level abstractions. It's not a hard rule, but more of a guideline. Papermill predates nbclient some so more of the flexibility and opinions grew there. Rather than moving the execution logic solely to papermill we wanted to have a less opinionated, simple execution library that can be inherited / called functionally from various applications so nbclient was made. All that being said, I'm not opposed to adding capabilities to nbclient. I was just pointing out that registering hooks fits more naturally with the abstractions in papermill than nbclient given the goal of the two libraries. So don't take my post as a rule to avoid improvements in nbclient please.
That's reasonable as well since nbclient is meant to be reused in more complex execution patterns. I would maybe think of having a new notebook execution function that takes a post-cell function and leave the rest as vanilla nbclient. We're unlikely to rework the execution contracts in nbclient before 1.0 and will continue to support methods we expose as best as we can. @davidbrochart uses nbclient in other libraries, so he might have a different flavor on how he views the libraries. Hope that clarifies some |
It clarifies a lot, thanks. NBConvert predates papermill (I think), and I think that's I mixed things up a bit I'll rework this to add hooks then. Thanks for the information. |
Yeah nbconvert (where nbclient came from) does predate papermill, though it's execution library was not being maintained for a while there. |
Anyway, I've added 4 basic hooks, which IMO expose very useful information. What other hooks do you think should exist, do you think this is being done the right way? Would love some opinions. |
@MSeal @davidbrochart any thoughts/comments about this? |
nbclient/util.py
Outdated
future = hook(*args) | ||
else: | ||
loop = asyncio.get_event_loop() | ||
future = loop.run_in_executor(None, hook, *args) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why not just call hook(*args)
here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think it'd be preferable to use kwargs so it's more forward compatible?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I don't want to block execution, which calling hook(*args) will do, so I'm executing it on a threadpool executor
Yeah good call on the kwargs.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
OK just remembered why I went with args..
run_in_executor only takes args, not kwargs....
I suppose it's worth it to use functools.partial for this though
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Seems simple enough overall. I'll think on if we want to skip traitlets for this or leave them, but for now the minor comments addressed and some clearer doc strings on the hook options and I'd be fine with a merge.
nbclient/client.py
Outdated
@@ -223,6 +223,35 @@ class NotebookClient(LoggingConfigurable): | |||
|
|||
kernel_manager_class = Type(config=True, help='The kernel manager class to use.') | |||
|
|||
on_kernel_create = Any( |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This might be deceptive to users, because it's only run when nbclient makes a kernelmanager and not when a kernel is created. The kernel creation aspect is fairly abstracted away from nbclient so I would instead make a notebook_start hook after the kernel setup is completed if you're going for pre-cell execution hooks.
nbclient/client.py
Outdated
help="""A callable which executes when the kernel is created.""", | ||
).tag(config=True) | ||
|
||
on_cell_start = Any( |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
These are somewhat awkward as Any
traitlets :/ Until we decide to move off them I guess this is how it'd be
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
agreed...
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Should we use a Callable
trait? It should also be typed as t.Callable
.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is there a Callable trait? I couldn't find one, and I copied the Any from the timeout func.
IMO, typing on traits is a bit redundant (except in this sort of case)
It could be typed like t.Optional[t.Callable]
since it can be None
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You're right, it looks like Callable
has been added to traitlets
in the past but then removed. I still think static typing traits is valuable, because it can catch bugs before runtime. You're right, it should be t.Optional[t.Callable]
.
nbclient/util.py
Outdated
future = hook(*args) | ||
else: | ||
loop = asyncio.get_event_loop() | ||
future = loop.run_in_executor(None, hook, *args) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think it'd be preferable to use kwargs so it's more forward compatible?
nbclient/client.py
Outdated
if self.force_raise_errors or not cell_allows_errors: | ||
if (exec_reply is not None) and exec_reply['content']['status'] == 'error': | ||
if (exec_reply is not None) and exec_reply['content']['status'] == 'error': | ||
run_hook(self.on_cell_error, cell, cell_index) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do you want this hook for any error (including ignored ones)? It might require that it be specified that suppressed errors would trigger the error handling hook. The caller below may not know if it's a suppressed error or not
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah I explicitly did it this way since people wouldn't use the hook if they were also suppressing errors. What kind of errors make it here that wouldn't be suppressed?
@MSeal thanks for the input and review. I've addressed most of the things. Not too sure what was to be made clearer in the docstring, but I've given it a stab. I've no idea why this is breaking tests... |
@golf-player probably good for now. Let's fix those two whitespace issues causing the linter tests to fail and I think we can merge. |
@MSeal fixed the trailing whitespace issue (not sure what I was doing when tox told me it was a problem....) |
also fixed conflicts and added typing to the function in util.py |
@golf-player that's great, could you add a test? |
Yeah I'll do some tests sometime this week. |
This will enable tracking of execution process without subclassing the way papermill does.
Hey guys, +1 for this PR 👍. In jupyter-book/jupyter-book#833 (comment) we were discussing about logic for skipping cell execution, e.g. if the cell contains a certain metadata tag. response = run_hook(self.on_cell_start, cell=cell, cell_index=cell_index)
if response is False:
self.log.debug("Skipping cell execution due to hook response %s", cell_index)
return cell (although the current async nature of the hook I guess makes it trickier) |
default_value=None, | ||
allow_none=True, | ||
help=dedent(""" | ||
A callable which executes before a cell is executed. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This may be just my ignorance on async, but is this sentence technically true?
It looks like in run_hook
you are enforcing these functions to be asynchronous, with no await, meaning that although they start execution before the cell is executed, they may not actually finish before the cell is executed?
Also, if this is the case, is it wise to be parsing a non-copy of the thread unsafe cell
object to the hook?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I believe you're right. There's no guarantee it executes prior to the cell being executed, which means potentially people could expect mutating the cell or something in the hook would occur prior to the actual execution of the cell. Thanks for the catch there.
Do you have any suggestions on how to handle this?
I mostly wanted this feature so I could do something like a live indicator of what cell was running at a given time (which wouldn't be a problem), so this didn't occur to me. And for the same reason, I didn't want to block the execution of a cell with a hook. Maybe an option to make the hook block? Or perhaps making the hook a coro means it gets executed as a Task, and otherwise it blocks?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Heya, yeh I think the hook calls should always be await
'ed. If you still wanted to add this non-blocking type behaviour I would say add it within your hook function, rather than it being intrinsically within nbclient (although personally I wouldn't advise it, because it feels like it could possibly result in a "mess" of task completion timings)
It also feels like maybe you should just specify that all hook functions should be Awaitable, rather than having this async wrapping behaviour that again is not made super clear to users from the traitlet
Sorry, I've been generally non-productive the last month or so and let this thing slip. I'll get it up to date with master and hopefully finish things off this coming week. Assuming I'm able to address the new comments |
Hi @golf-player, do you still plan to work on this PR? We were discussing in jupyter/nbconvert#1380 and we think that it would be helpful. |
Hi @golf-player @davidbrochart - I'm interested in this enhancement as well. If I'm following the conversation correctly, it sounds like the only things missing is a rebase and tests? If so, I'd be willing to continue this. Not too sure what the etiquette is on continuing another person's PR as well, so any guidance on that would be appreciated. |
Hi @devintang3, if you want to continue that work, great! Rebasing and adding tests would be a good start. |
I'd find it useful to have the client expose some information about what's going on during the execution.
I'm particularly interested in knowing what's happening in the execution at the current time, and also which cell is being run.
This is very, very rough, and just to give you the gist of what I'm looking for. Let me know if y'all would be interested in such a feature and I'll make it less rough and add tests and stuff.
Please let me know what you think. Maybe rather than doing it like this, exposing hooks for users could be the way to go?