PHP RFC: Server-Side Request and Response Objects
- Version: 2.2
- Date: 2020-03-17
- Author: Paul M. Jones, firstname.lastname@example.org
- Status: Declined
- First Published at: http://wiki.php.net/rfc/request_response
This RFC proposes an object-oriented approach around request and response functionality already existing in PHP, in order to reduce the global mutable state problems that come with superglobals and the various response-related functions.
The SQLite “about” page says, “Think of SQLite not as a replacement for Oracle but as a replacement for fopen().” https://www.sqlite.org/about.html
Likewise, think of this RFC not as a replacement for HttpFoundation or PSR-7, or as a model of HTTP messages, but as an object-oriented alternative to superglobals, header(), setcookie(), setrawcookie(), and so on.
This RFC proposes an extension to declare three new classes and one interface in the root namespace:
- SapiRequest, composed of immutable read-only copies of PHP superglobals, and some other commonly-used values parsed out from those superglobals
- SapiResponse and SapiResponseInterface, a buffer for response-related PHP functions
- SapiResponseSender, to emit the SapiResponse using PHP functions
The full README, working code, and all tests are available at https://github.com/pmjones/ext-request, in the 2.x branch.
An earlier version of the extension is in the 1.x branch of the same repository; differences from that version are discussed later this proposal.
Instead of the superglobal ... ... use SapiRequest: --------------------------------------- --------------------------------------- $_COOKIE $request->cookie $_GET $request->query $_GET['key'] ?? 'default' $request->query['key'] ?? 'default' $_FILES $request->files $_POST $request->input $_SERVER $request->server $_SERVER['REQUEST_METHOD'] $request->method $_SERVER['HTTP_HEADER_NAME'] $request->headers['header-name'] $_SERVER['CONTENT_LENGTH'] $request->contentLength $_SERVER['HTTP_CONTENT_MD5'] $request->contentMd5 $_SERVER['PHP_AUTH_PW'] $request->authPw $_SERVER['PHP_AUTH_TYPE'] $request->authType $_SERVER['PHP_AUTH_USER'] $request->authUser Instead of parsing ... ... use SapiRequest: --------------------------------------- --------------------------------------- $_FILES to look more like $_POST $request->uploads $_SERVER['CONTENT_TYPE'] $request->contentType and $request->contentCharset $_SERVER['HTTP_ACCEPT'] $request->accept $_SERVER['HTTP_ACCEPT_CHARSET'] $request->acceptCharset $_SERVER['HTTP_ACCEPT_ENCODING'] $request->acceptEncoding $_SERVER['HTTP_ACCEPT_LANGUAGE'] $request->acceptLanguage $_SERVER['HTTP_X_HTTP_METHOD_OVERRIDE'] $request->method $_SERVER['PHP_AUTH_DIGEST'] $request->authDigest Instead of emitting ... ... buffer with SapiResponse: --------------------------------------- --------------------------------------- header('HTTP/1.1', true, 200) $response->setVersion('1.1'); $response->setCode(200); header('Foo: bar', true); $response->setHeader('Foo', 'bar'); header('Foo: bar', false); $response->addHeader('Foo', 'bar'); setcookie('foo', 'bar'); $response->setCookie('foo', 'bar'); setrawcookie('foo', 'bar'); $response->setRawCookie('foo', 'bar'); echo $content; $response->setContent($content); Instead of sending with ... ... send with SapiResponseSender: --------------------------------------- --------------------------------------- echo, header(), setcookie(), etc. $responseSender->send($response);
There is more: please see the docs at https://github.com/pmjones/ext-request/blob/2.x/README.md for more information.
Criticism and Objections
Why Do This In PHP Itself?
For a language as closely related to the web as it is, PHP has lacked core server-side request and response objects for its entire existence. It has date objects, database objects, XML objects, and other object-oriented extensions, but none for server-side request and response objects. This proposal fills that gap, covering a set of common functionality that has until now been available only in userland.
Further, truly read-only objects in PHP userland are difficult if not impossible to achieve, especially when you take immutability of values into account. Working within an extension is the surest way of doing it, a la the request object offered here.
What About Existing Userland Projects?
I would prefer to discuss this proposal on its own merits, and thereby stay away from what might appear to be negative commentary on other projects. Nonetheless, prior critics of this proposal have demanded comparison between it and pre-existing projects.
With all that in mind, I will present limited comparisons to two other major projects, hopefully hitting the high points without sounding overly-negative. If further comparison is desired, I will attempt to provide it.
HttpFoundation provides a very wide range of functionality, as evidenced by its documentation at https://symfony.com/doc/current/components/http_foundation.html. It does so at a cost of necessarily greater complexity and higher code volume.
As it happens, this proposal turns out to mimic a reduced subset of HttpFoundation functionality. The same subset is common to many userland implementations: Aura, Cake, CodeIgniter, Horde, Jooma, Klein, Lithium, MediaWiki, Nette, Phalcon, Yaf, Yii, and Zend/Laminas (among others). That subset is:
- a way to read the request-related superglobals such as $_GET, $_POST, etc. from an object; and,
- a way to set headers and content into an object so they can be inspected and modified before sending.
So, this proposal is in some ways a distillation and summary of widely desired functionality in userland.
Back to HttpFoundation specifically, an example of setting cookies provides an illustrative comparison:
use Symfony\Component\HttpFoundation\Cookie; $response->headers->setCookie(Cookie::create('foo', 'bar'));
This is not so terrible, though it does additionally involve the HeaderBag and Cookie classes. Whereas with this proposal, again intended only to wrap existing PHP functionality, the code would be:
HttpFoundation request and response objects are both fully mutable. In contrast, this proposal offers a request object with read-only properties; the class can be extended to add mutable or immutable properties. The response object offered here is fully mutable, though its methods are marked as final; this leaves the response object open for extension by frameworks and libraries, but closed for modification to its core functions.
(Full disclosure: I was one of the sponsors on PSR-7.)
PSR-7 is a newer competitor to Symfony HttpFoundation. The PSR-7 interoperability interfaces, and their various competing implemementations (each with their own idiosyncrasies and additions), attempt to model HTTP messages, both for use by an HTTP client (as in Guzzle) and by server-side applications (as in Laminas).
Using PSR-7 for server-side requests and responses can be challenging. For good or bad, the specification defines a way of working that is very different from implementations pre-existing it. A number of followup PSRs have been created to relieve these issues, as have other userland helper packages.
Using the same example as with HttpFoundation above, setting a cookie in PSR-7 is no simple task. PSR-7 only provides a way to set headers, which generally means a helper library is necessary to set cookies properly. For examples of such libraries, see https://github.com/dflydev/dflydev-fig-cookies and https://github.com/hansott/psr7-cookies, among others.
Using the FigCookies project, this is how to set a cookie into a PSR-7 Response:
use Dflydev\FigCookies\SetCookies; use Dflydev\FigCookies\SetCookie; $response = $response->withAddedHeader( SetCookies::SET_COOKIE_HEADER, SetCookie::create('foo', 'bar') );
As such, PSR-7 is not “batteries included”: a 3rd-party helper library is needed for cookie work.
In contrast, this proposal offers something much more straightforward: not to model HTTP messages, but to wrap existing PHP functionality in object properties and methods. It is as about as easy as using PHP itself.
$response->setCookie('foo', 'bar', $options);
In a polar opposite of HttpFoundation, PSR-7 specifies immutability of both the request and the response. It turns out this is only imperfectly achievable for multiple reasons, not least of which is that the streams used for content are themselves mutable.
To reiterate, this proposal offers read-only properties on the request with consistent and reliable immutability of those values. The response object remains mutable.
Userland Availability, Comparability, and Ecosystem
(Copied, with light editing, from https://externals.io/message/108436#108493.)
One common objection, with variations, has been: “There is a wider userland ecosystem that already performs the proposed functions, with more capabilities, and with potentially hundreds of thousands of implementations already in place. Does the proposal add capabilities which do not or cannot exist in userland? If not, then leave it to userland.”
The proposal authors recognize and understand the sentiment. The following counterargument, in relation to previous PHP extensions, is presented in return.
When ext/pdo was added to core, there was already a “wider ecosystem that already performs these functions, with more capabilities, and with potentially hundreds of thousands of implementations already in place.” Some of those implementations at the time included AdoDB, Metabase, MDB, PEAR_DB, and many more.
PDO did not “add capabilities which do not or cannot exist in userland”. (The proposal authors grant that FETCH_INTO_OBJECT setting properties directly without using the constructor was not possible in userland, but that's an exception that tests the rule.) Indeed, PDO had a relatively reduced feature set in comparison to some of those userland libraries, especially AdoDB.
And yet, PDO has turned out to be of great benefit, because it brought together features into core that (figuratively speaking) everybody needed and was rewriting in userland over and over.
PDO is the strongest example here, but depending on how you count, there are 2-3 other extensions that also serve: ext/date, ext/phar, and (reaching back to antiquity) ext/session.
So, there is a long history of widely-needed userland functionality being brought into core. This proposal is a pretty tame example of doing so; as presented, it is very similar to the way PHP itself already does things, just wrapped in object properties and methods, and is very similar to how things are being done across a wide swath of userland.
Now, it is possible that the above objection should have prevented PDO (et al.) from going into core. If that is the case, and (in hindsight) it was a mistake to allow them, then consistency alone makes the objection valid here as well.
However, if (in hindsight) it was not a mistake to allow those extensions, then the objection is not an especially strong argument against this RFC. That's not to say “because PDO was allowed into core, this RFC must therefore be allowed into core” but to say “that objection alone would not have been a barrier to PDO, so it alone should not be a barrier to this RFC”.
Other Questions And Comments
Q: The proposal compares and contrasts with HttpFoundation and the various PSR-7 implementations; how does it compare to other projects?
A: See this message for a starting point: https://externals.io/message/108436#108889. In short, the proposed functionality is representative of functionality common to the dozen-plus researched projects.
Q: Are these global single-instance objects?
A: No, you can create as many instances as you like, in whatever scopes you like.
Q: Do these objects replace the superglobals?
Q: Do these objects deal with $_SESSION and the session functions?
A: No; it is explicitly out of scope for this RFC.
Q: Does SapiRequest hold references to the superglobals, or copies?
A: Copies, made at instantiation time. Changes to `$_GET` after the SapiRequest is instantiated will not be reflected in the existing instance.
Q: Since the $query, $post etc. properties are the same as $_GET and $_POST, does that mean they retain the same name mangling scheme?
A: They do; that is, SapiRequest uses whatever is passed into it at construction time. If PHP changes its name mangling, or if different array values are passed in, SapiRequest will use those instead.
Q: Readonly properties are unusual for PHP.
A: Granted, though not unheard of. PdoStatement::$queryString is one precedent here. Further, of the researched userland projects, more than half of them present the relevant superglobals as nominally readonly in the public scope, making readonly access a reasonable choice here.
Q: Does this has any performance impact?
A: Compared to userland, probably greater performance, but the scope is so small that I expect little end-to-end impact on applications as a whole.
Q: Why is SapiRequest readonly, and SapiResponse mutable?
A: It makes sense that you would not want to change what you have received as a request; however, as you are in charge of creating the response, modifying it as needed seems reasonable. Further, the “readonly request with mutable response” matches half or more of the researched userland projects.
Q: Why is SapiRequest composed only of properties, and SapiResponse composed only of methods?
A: It's an outgrowth of an asymmetry that already exists in PHP: $_GET, $_POST, et al. are properties representing the request, whereas header(), setcookie(), et al. are all functions for sending a response.
Q: Why not write (PSR-7|HttpFoundation|OtherImplementation) in C, instead of your own version?
A: This is not “my own version.” This is an OO-approach to what PHP itself already does; it is representative of PHP's way of doing things, not “my” way of doing things.
Q: Does it support HTTP/2?
A: It supports HTTP/2 exactly as much as PHP itself does.
Q: Does it support async?
A: Async is not in scope for the proposed API.
Q: What would a migration path look like?
A: Something like the one outlined in the later portion of this message: https://externals.io/message/108436#108893
Changes From The 1.x Version
Based on user feedback over the past couple of years, this proposal differs from the earlier 1.x version in the following substantial ways:
- The “Server” prefix on the class names has been changed to “Sapi”.
- Some users objected on principle to the SapiRequest constructing itself using the superglobals internally. As a result, the constructor now requires a single array parameter; the corresponding argument is typically $GLOBALS but can be any array that mimics the $GLOBALS structure.
- The SapiRequest object no longer has the immutable application-related functionality represented by withInput(), withParams(), withUrl(), and their sibling methods. Some users felt this functionality was better left to application-specific implementations; other users merely did not need them and were happy to ignore them. This means SapiRequest is now only a public read-only set of properties with immutable values, while still being extensible in userland for application concerns if desired.
- The SapiResponse object no longer has setContent() convenience methods such as setContentJson() and setContentDownload(). Users found these less convenient than anticipated, and preferred to add their own application-specific convenience methods.
- SapiResponse no longer has a self-sending capability. It was noted that to customize sending logic, you needed a custom SapiResponse object. As a result, the sending logic has been extracted to a SapiResponseSender class.
- To address some concerns from an earlier round of discussion, all SapiResponse properties are now private, and all its methods are now final, though the class itself is not. This keeps the class open for extension but closed for modification.
- SapiResponse::setHeader() and addHeader() methods no longer convert array values to CSV header strings; this functionality was so rarely used as to be unnecessary. Removing it brings these methods back in line with the PHP header() signature. Likewise, the date() helper method is similarly removed. These helper methods, if ever needed, are easily added to userland implementations.
In all, these removals and changes bring the proposal much closer to PHP as-it-is.
1. Should these classes go into an existing extension, rather than one of their own? Or should they go into “core” proper?
Backward Incompatible Changes
Userland code that declares classes named SapiRequest, SapiResponse, or SapiReponseSender will need to rename those classes.
Proposed PHP Version(s)
Next PHP 7.x or 8.0.
To Existing Extensions
It would be convenient if the php_head_parse_cookie_options_array() function in ext/standard/head.c was published in head.h, so that the logic there would not need to be duplicated in php_request.c. Doing so would make it easier to keep the logic in sync, but is not a requirement.
None at this time.
Unaffected PHP Functionality
The remainder of PHP should remain unaffected.
This extension acts as an object-oriented wrapper around existing PHP request and response functionality; as the scope of that PHP functionality expands, this extension should expand with it.
Proposed Voting Choices
For or against the proposal.
Patches and Tests
The C code for the extension and tests are at https://github.com/pmjones/ext-request in the 2.x branch.
After the project is implemented, this section should contain
- the version(s) it was merged to
- a link to the git commit(s)
- a link to the PHP manual entry for the feature
https://externals.io/message/108436 (discussion thread)
https://externals.io/message/109161 (voting thread)
Add filter_input integration to SapiRequest.
Add .ini setting(s) to disable superglobals, and/or warn on their use.
Add .ini setting(s) to disable response-related functions, and/or warn on their use.
Expand the number of classes provided, to allow for various SapiRequest-related value objects.
Provide builder and locking methods for SapiRequest.
Make the SapiRequest properties mutable.
Add a SapiResponse::addContent() method.
Embed the PHP multipart/form-data and application/x-www-url-ncoded parsing mechanisms into SapiRequest, possibly exposing them wider use.