Open KoenLav opened 8 years ago
It is important that unbind
isn't called twice when bind
is. It must have something to do with some detail in onRefreshed
, onDestroyed
or onInvalidate
. One of those hooks must fail to fire in your special situation.
I'm pretty sure unbind is 'working' (partially), because the .off() is being called (and the listeners are, if only very briefly, detached).
However I think that because the accompanying Nexus is not removed more than one listener is added when bind is called.
Right. So the view's onViewReady
hook calls the bind
function of both the old and the new nexus. That makes sense. That means unbind
ought to somehow unregister bind
from onViewReady
(after which the nexus will be garbage collected instead of bound again).
The weird thing is that this problem only occurs sometimes. There must be something special going on with onViewReady
that differs in your case.
In List.remove I do console.log(this, items) when it was called from unbind using this.view[ViewModel.nexusesKey].remove(this)
Then I get temp1 and temp2 and use temp1.indexOf(temp2) and get -1.
Because: temp1 contains two Nexuses with selector vm-bind-id 258 and 287, but temp2 contains one Nexus with selector vm-bind-id 128.
I'm fairly sure all these issues are related to the vm-bind-id being changed, for some reason.
The vm-bind-id
attribute should change every time the view is refreshed – it is expected behavior. It's the only way the {{bind}}
helper can be used in that special way that we use it.
I see, but because it is changed before/during unbind is running the object passed to remove does not match any of the objects in the list (I think).
Other places that would be interesting to inspect are view._callbacks
and view._domrange.attachedCallbacks
.
I've continuously profiled the usage examples for memory leaks and made sure there wasn't any, so I'm surprised that the library is bleeding nexuses in your case.
But I guess I haven't profiled any cases where a view is refreshed the way that it is in yours.
It would be helpful with a reduced test case that demonstrates this bug, but I know it isn't always a simple thing to do.
I don't have time to create a test-case currently (can only spend a couple of minutes at a time), but maybe tonight or tomorrow :)
I will keep working on trying to debug it though.
:+1:
I've pinpointed the issue to the Tracker.afterFlush function in the onReady function; it fires twice.
Nice find!
Which makes perfect sense, since the onReady function also gets called twice, apparently.
I may not be able to look into it more until tomorrow, but I'm sure we will be able to fix this within a very short timeframe.
Great to hear, I've got a few more minutes today ;)
Continuing:
Which makes perfect sense again, since the element has had two nexuses attached to it. I'm not sure whether Tracker.afterFlush is the 'right' method to use here.
Anyhow I'm confident I got closer to the source of the issue, so I'm going to all changes and restart at onReady/Tracker.afterFlush.
Ok, I noticed something:
We have a list of 30 elements experiencing this issue, but sometimes some of the elements will result in a double bind whereas others will not... "Race condition-like" behavior.
It seems some things are just not completed in the expected order, but which things I don't know yet.
Also when I reconnect a second time some elements' viewmodel functions will fire twice whereas others will fire thrice.
Monkey patched the bug by adding
if (!view.afterFlushSet) {
Tracker.afterFlush(callback);
view.afterFlushSet = true;
}
On line 63 in base.js, but have yet to find the cause of the Nexuses not being removed.
Never mind, this causes other issues for the elements which were not problematic earlier...
Ok, what happens at the remove stage:
List of Nexus.selectors: [vm-bind-id='250'] [vm-bind-id='282']
Resulting in the new list of Nexus.selectors: [vm-bind-id='250']
These two Nexuses together cause the double triggering of the viewmodel functions.
Ok, I further pinpointed the issue: unbinding works as expected in our scenario, it is the binding however which gets executed twice and messes things up (probably for the same reason Tracker.afterFlush gets called twice).
Would it help to know that these elements are rendered within a loop within a loop?
{{#each category}}
{{#each products}}
{{> productTemplate}}
{{/each}}
{{/each}}
Simplified representation (these are actually split over two templates).
I think it's an interesting perspective that we might be dealing with a race condition. Is it always the exact same elements that get the duplicate bindings?
Yes, but as I said it's a list of 30-something elements and there are times when elements 23 and 24 (for instance) are unaffected but the others are affected (which supports the view of 'race condition-like' circumstances).
If I understand you correctly, it is not always the exact same set of elements that gets affected?
It's 30-something elements which all make use of the same template, most of the times all of them are affected but on rare occasions some of them are unaffected.
So you are right in saying that not always all of the elements in the set are affected, but the phenomena is always limited to a subset of this specific set of elements.
My guess is that Meteor.reconnect()
triggers a Tracker flush, which coincides with some or all elements being in a temporary state where the onReady
or other event are supposed to happen before (or possibly after – I forget) an already scheduled flush.
The solution would be something along the lines of what you tried earlier. Hmm...
If you set a breakpoint on this line and look in the call stack, does that part of onReady
ever get called directly by the Nexus
constructor, or is it always inside a call to autorun
(= irrelevant)?
Following the code down the rabbit hole:
For now I ended up at the bindHelper function of the viewModel which, apparently, gets called twice every time reconnect is called (for these elements).
However right after bindHelper is called for the second time the onDestroyed which resulted from the first bindHelper is called. The onDestroyed calls from the first bindHelper calls unbind on the element, but removing the Nexus returns false (I guess because the second bindHelper has already fired and changed the vm-bind-id of the element, causing a mismatch somewhere down the road).
I'm not sure why these elements are so rapidly re-rendered (possibly because GroundDB causes some computation to invalidate twice, rather than just once, on a reconnect), but I guess this could occur in other scenario's as well.
So I think we should make sure that when binding a helper any previously existing binding on the same element originating from the same Blaze helper should be properly removed.
The order of events:
bindHelper ONE viewmodel.js:512 newNexus nexus.js:114 ===== nexus.js:103 onDestroyed nexus.js:300 unbind nexus.js:313 false <== cannot remove nexus for some reason base.js:66 afterFlush nexus.js:255 bind <== we are binding here viewmodel.js:483 click: addProduct viewmodel.js:510 =====
bindHelper TWO viewmodel.js:512 newNexus nexus.js:114 ===== base.js:66 afterFlush nexus.js:255 bind <== we are ALSO binding here
Looking back at my previous reply I'm not sure whether unbind ==> false is the issue, it might rather be the fact that afterFlush is called twice resulting in two binds.
Working on that breakpoint now
For sure, it could happen in other scenarios – thanks again for taking the time to debug this. I'll look into the reason why do_unbind
evaluates to false.
The {{bind}}
helper being triggered twice is expected behavior – dynamic attributes are re-run when anything about the element changes, which is the reason for many of the tricky maneuvers employed by the library.
Just to be sure, you are using the latest version 1.0.2
, right?
I'm not really used to working with the stack directly, but the when I set a breakpoint on that line and look in the Chrome DevTools the second line says "nexus.js:111" which is this.onReady(this.bind) in the Nexus constructor.
The false I mention there is not the do_unbind, it is the result of this.view[ViewModel.nexusesKey].remove(this);
The nexuses
list on the view is only for debugging and doesn't affect the bindings.
Do you think I should focus on why this.view[ViewModel.nexusesKey].remove(this) returns false or rather on why the bindHelper is called twice.
There are two differences between elements that don't have the erronous behavior and those who do. For elements that have the bug:
I see, if that is the case then Nexus.remove(this) a few lines earlier would probably also return false. I'll look into that.
It's weird if it returns false
, but those two lines have no bearing on the bindings themselves, they only track what is currently on the page for debugging purposes.
The bind helper being called twice is expected behavior and only causes problems when the previous binding isn't unbound properly before each subsequent call.
Nexus.remove(this) also returns false.
But I see your point.
I do still think the afterFlush ==> bind being called twice is the culprit. Every time reconnect is called the previous binding is removed, but two new ones are added due to the afterFlush being triggered twice.
afterFlush should never trigger a bind if the previous bind has not yet been removed, I guess.
Actually I'm 100% sure this is the issue.
Even though Nexus.remove(this) returns false, before that the eventListener is removed so that should be fine (in terms of functionality, not in terms of memory leaks).
However because afterFlush is set twice (because it is bound in two different computations, one for the first triggering of bindHelper and another for the second) it is executed twice, but this shouldn't happen.
I'm still not sure whether afterFlush is the right 'tool for the job' (but I'm not sufficiently equipped to think of a better alternative).
Here's another idea: unbind
accidentally runs before bind
does, which would explain why Nexus.remove(this)
returns false – it can't be removed from the global list, because it hasn't been added to it yet (which happens in bind
).
The first unbind
ought to run after the first bind
, of course.
Don't pay too much attention to the multiple calls to bind
/unbind
, by the way – it's the only way that the library can be guaranteed to work in all situations and what makes it especially robust, in all modesty. Except in your case, of course ;)
I wonder whether view.onViewDestroyed
sometimes runs before view.onViewReady
? If the view is destroyed immediately, is the unbind
callback run right away when registered? In that case, the value of view.isRendered
and view.isDestroyed
is important...
Still wish I had a test case ;)
To muddle the image, list.add
and list.remove
will both trigger a Tracker flush.
Have we established exactly which hook is responsible for running unbind
? Is it onRefreshed
, onDestroyed
, or onInvalidate
?
onDestroyed
I am using these console.logs:
this.onRefreshed(function() {
if (Meteor.checkElement(view)) {
console.log("onRefreshed")
}
that.unbind()
});
// Unbind element on view destroyed
this.onDestroyed(function() {
if (Meteor.checkElement(view)) {
console.log("onDestroyed")
}
that.unbind()
});
// Unbind element on computation invalidation
this.onInvalidate(function() {
if (Meteor.checkElement(view)) {
console.log("onDestroyed")
}
that.unbind(true)
});
In the Nexus constructor.
Hi Kristian,
We are happily using your package for a while now. But since a while it seems that events are binded to HTML nodes twice after Meteor reconnects.
Have you seen this problem before?
Where is the code located which binds the event (within your package)? Maybe I can take a look ;)