I am not sure if the BC is allowed or if we should add an argument to control the behavior and deprecate (but this kinda defeats this change as goal seems to make this more developer friendly)

We shouldn't break BC, so a constructor argument to enable this behavior would be needed.
Then there should be a similar argument toBrowserKit\Client so that this could be also enabled there. Before doing so, do we all agree returningnull is the best for DX and we should move out of the exception?

Here is a better proposal@nicolas-grekas to avoid BC as you suggested (way better), not sure about the new argument name.

Now, for theBrowserKit, new PR or two commits same PR ?

We still have a bunch of other methods which fail for an empty node list.

And btw, making return types nullable will not save you for doing a check to handle that case. Instead of checking whether the crawler is empty or no, you will have to check whether the value isnull.

and regarding BC, this is still changing the contract of the class. Some code receiving a Crawler as argument will not control this boolean, and so would have to be updated to support the new signature as well. So this approach is still a BC break.

@stof looking at the original issue, it could also be empty, let's say an empty string, in this way it keeps the signature.
Do you have any suggestions? Or do you propose to drop this change?

Well, returning the empty string is possible for these 2 methods. But then, we still cannot remove the exception from other methods (the ones returning an object for instance)

A test fails but seems to be a travis problem with composer?

Now, for the BrowserKit, new PR or two commits same PR ?

@respinoza how would you leverage this new argument? Do you create your crawlers manually?

Withtext() andhtml() methods returning an empty string how can I tell the difference between an empty text and element not found?

@jakzal thecount() method still return the matches, so0 for not found.

@nicolas-grekas we will have to pass it to the BrowserKit Client constructor and then pass it around inhttps://github.com/symfony/browser-kit/blob/68f233bc3a0fb6b64f48ff6318038087b49a1639/Client.php#L516`

Currently there is no way to inject a DomCrawler instance. This can also be solved by improving the DomCrawler/BrowserKit as follows

• Add method to DomCrawler to create a new instance (similar like the currentcreateSubCrawlermethod)
• Add setter to pass the DomCrawler to the BrowserKit client
• Get a fresh instance if we set a custom DomCrawler instance or create a new like we do right now

Currently there is no way to inject a DomCrawler instance

so let's allow it via a new constructor arg to BrowserKit's Client?
should be in the same PR to make it complete (including the DI configuration)

so let's allow it via a new constructor arg to BrowserKit's Client?

hum, looks like a crawlerfactory would be needed, so now I get your proposal.
I'd still prefer passing this factory as a constructor arg instead of a setter.

@nicolas-grekas having a factory will be way better. I will get on that.

I'm not sure passing a Crawler factory just to control this setting (which looks transitional) makes sense.

Injecting a Crawler indeed does not make sense, as the Crawler holds some content, so it does not make sense to inject one that the Client would not know what is already in it.

@stof the current code in BrowserKit creates a new instance or null if the class doesn't exist. The factory will be called instead, behavior is the same (new instance on method call), we will just show to return a configured instance.
Sub crawlers are created as usual.
We won't reuse the crawler.
I don't see a problem

Well, the question is whether we need to create a concept of factory when theonly meaningful configuration to be customized by the extension point is this boolean.

Currently, this PR makes that setting a permanent option. It does not deprecate the old behavior. Is it intended ? Variable signatures based on an option are harder to use (as any code not creating the crawler could not rely on whether an exception could be thrown or no for empty crawlers, and so should keep handling the case throwing exceptions too, defeating this PR).
If this becomes a deprecated API with a migration path, making people replace the whole instantiation of the crawler is too much:

• it makes the migration harder as they have to duplicate the core logic in 99% of cases
• it introduces an extension point that we then have to maintain even if we don't need it for the BC layer. We cannot remove it in 5.0 (as we cannot deprecate it in 4.x if using it is mandatory to avoid another deprecation).

Thus, if the goal is to deprecate the old behavior, I think a constructor argument is the wrong way to opt-in for the new behavior: the code needing to deal with the new behavior (and so knowing whether it can opt-in or no) may not be the same than the code instantiating the crawler (and when using BrowserKit, that would mean the code instantiating the client, to configure it to configure the crawler creation...).

I don't believe we want to get rid of the old behavior so there won't be any deprecation.

I agree with your point on the variable signature. We could always have two new methods and get rid of the parameter.

(No idea how to make those methods)

@jakzal creating a decorator will be a pain:

• adding any method is a BC break, unless we also provide a base class delegating all methods in DomCrawler itself and consider that a decorator must extend our own base class to be subject to BC. Otherwise, the newly introduced method would get called on the parent class rather than on the inner object
• things become even more complex due to returning sub-crawlers from many methods
• changing the signature in a decorator class is still breaking the contract for any code consuming a Crawler.

At that point, it is much better to have this text extraction utility in your own code rather than in a crawler decorator.

And regarding the discussion about BC, a configurable signature is a bigger BC break than a variable signature based on a new optional argument. A variable signature creates a BC break for child class (and the migration path is clear, with the possibility to trigger a deprecation). A configurable signature creates a BC break for consuming code (and with no way to warn them) due to the signature change (as they might not be the one controlling the contrustor-based opt-in).

Option 1. is possible, using func_get_arg() to prevent any BC break. We already did so in other places.

I added the argument to the function throughfunc_args as discussed for option one.
Looking at the related issues, I guess it will be a good idea to extend this to theattr() method?
That one already has an argument, should still do afunc_args and check for the second one?

Also again not sure about the parameter name (naming things is hard), ideas welcome 😅

@respinoza what about$throw = true :) then we can also apply it onparents() etc. maybe, returning an empty list in case throw=false.

Passing true for every method is also a hectic approach. It would have been better if we could somehow transfer this approach from the Crawler constructor.

A default value, rather than a boolean flag, would be more flexible.

I would say no to a default value as we don't want to change the signature (return single), we could check it is a string but in practice is someone going to have another value than empty string?

Maybe the idea of having new methods to handle this might be better.

Any more ideas?

Yes, two :)

1. name the new argument "$default" and return it if provided (and throw when not provided)
2. do nothing and keep things as is

@nicolas-grekas sorry for the delay, changes have been done and hopefully it can be included for 4.3

Thank you@respinoza.

I'm following after#28614, in the end this has not been applied toattr() which still throws the exception.
Apologies, It was not available in my local version, but it has been added in 7.0.0-RC1
94af00e