On Mon, 9 Mar 2026 at 13:37, Gina P. Banyard <[email protected]> wrote: > > On Monday, 2 March 2026 at 19:54, Christian Schneider <[email protected]> > wrote: > > > Am 02.03.2026 um 20:12 schrieb Tim Düsterhus <[email protected]>: > > > On 3/2/26 14:49, Christian Schneider wrote: > > >> Playing my favourite broken record: > > >> Can we please state that additions of Exceptions should (in most cases) > > >> go through an E_WARNING phase to allow a time window to fix code before > > >> changing the behaviour? > > > > > > “Not passing invalid values” is perfectly backwards compatible. Folks can > > > just fix their code before upgrading their production deployment to the > > > new PHP version, e.g. by trying out the new PHP version in a staging > > > system or running CI for both the old and new PHP version. > > > > - Not everybody has access to a staging system, e.g. people running stuff > > on hosting services. > > - As a hoster I'd rather have a phase where my customers get warnings > > instead of errors, creates less emergency support load. > > > > > In practice an E_WARNING is no less breaking than going straight to an > > > Error, because: > > > 1. The common frameworks include error handlers that just convert any > > > warning and notice to an Exception. > > > > So in that sense there is also no advantage to NOT having a warning phase > > for those people. > > But people treating E_WARNING different from Exceptions (which is probably > > the exact people whose code breaks with an immediate Exception) do have a > > time window to fix things. > > > > > 2. The code is already broken, because it relies on unspecified behavior. > > > The error would just making the user aware that the code is very likely > > > not doing what it appears to be doing based on the input values passed to > > > a function. > > > > It can easily do something valid and ignore the extra bits (pun kind of > > intended), see > > mkdir("foo", 070777); > > which passes extra bits with are ignored but the application was behaving > > in a completely deterministic and valid way. > > > > > Going through an E_WARNING would add maintainer busywork and complicate > > > the php-src codebase for no real gain. > > > > We've been over this before: > > If people *really* feel that the additional burden to change the code twice > > then I'd be happy to volunteer providing a small helper function/macro to > > generate an E_WARNING and either (the more aggressive approach) switch to > > an exception once a certain PHP version is reached or (probably the more > > flexible way) issue a compile time warning/error informing the maintainer > > to switch the warning to an Exception. > > The most simplistic version of this is a FIXME comment annotated with a > > version number. Easy to grep, easy to trigger automatic alerts about once > > the specified version is reached, but it can also be something more > > sophisticated.. > > One way or the other could also be integrated into the CI/CD system. > > > > Our main disagreement is about the "no real gain" part as it IMHO targets > > the long tail of PHP code / developers out there not using full-fledged > > frameworks and dev environments or running legacy software on hosting > > services. > > > > I am very much in favor of making things easy for the code developers but I > > am of the strong believe that it can be done in a good way for developers. > > I don't see how going through an E_WARNING phase is helpful, rather I see it > as detrimental. > Foremost, what is the behaviour of introducing a warning? > Do we exit early and return false? > Or do we just warn and continue to use a possible default or nonsensical > value? > AFAIK every time a warning got introduced it followed the first approach, so > this doesn't seem to address the concern of giving developers more time. > However, if we do continue using the prior behaviour then we haven't solved > the concern in the slightest. > As it may be a warning for one PHP version and in the next PHP version the > extension supports a new flag which removes the warning for that value, > leading back to a silent BC break if the warning wasn't addressed. > > Another technical aspect is that warnings allows userland code to run and > change the state of the VM and extension in unexpected way (we have countless > fuzzying reports about this) an issue that does not exist with Exceptions. > > Then comes the topic about how long should it be a warning? Until the next > major? A single release cycle? I don't want warning promotion to become the > same exhausting discussion that deprecation duration already is. > > I have no idea how you handle hosting, but when I used shared hosting in the > past I would never have anyone tell me that my code was producing warnings or > notices. > And even the one website I manage for someone which is on OVH shared hosting > I can still select a PHP version (heck even downgrading to 5.4 for some > reason). > So I'm struggling to see how for the average end user this is impacting them? > And if you do just bump the PHP version for them, I'm not sure what's the > difference between having an Error immediately or producing warnings for X > years before throwing an Error and breaking > > Best regards, > > Gina P. Banyard
I am in favour of exempting the validation errors from requiring RFCs, and as Gina, I do not believe that warnings are useful. I don't like that warnings even exist in the PHP language. I can see the case that you don't want to break a running application by upgrading the PHP version, but it can happen regardless. Minor PHP releases can contain minor breaking changes and each time someone upgrades PHP, they should run their test suite and verify that the application still works the same as before. Whether we introduce a Warning or ValueError, it will be caught during the upgrade process and fixed. It may even help avoid nasty bugs if the application stops on invalid input; something which could go unnoticed if it were only a warning. I am, however, concerned about one thing. If we don't require RFCs there may be situations when a contentious validation is introduced. For example, many validations can be implemented in such a way that they don't cause additional performance loss, but let's say someone decides that validating the value provides more benefit than the performance cost. Without RFC, the community cannot share their feedback, and one person's opinion wins. But if the function is used with the correct values 99.99% of the time and it's used in the hot code, the performance cost isn't worth catching the accidental invalid value.
