Download the PHP package hryvinskyi/magento2-browser-cache-control without Composer

Magento 2 & Adobe Commerce Browser Cache Control

Gives the visitor's browser a short Cache-Control lifetime on full-page-cacheable storefront responses, from inside Magento — without shortening how long the shared cache keeps them.

Description

Magento never lets a browser cache a page, and it does so for a different reason in each caching application. Normally the fix is two lines of VCL:

This module is for the installations where that edit is not available: a managed Varnish whose VCL you do not control, or a stack with no Varnish at all, running Magento's built-in full page cache.

It never invents cacheability. A response qualifies only if Magento itself declared it full-page cacheable, or served it out of the full page cache. Everything else — checkout, customer account, anything marked uncacheable — keeps the headers Magento gave it, untouched.

What Magento does today, and why that shapes everything here

Four facts out of the 2.4.x source. They are the whole design rationale.

1. Response\Http::setPublicHeaders($ttl) emits Pragma: cache, Cache-Control: public, max-age=<ttl>, s-maxage=<ttl>, Expires: now+ttl. It is called for every cacheable page, and also by the ESI block endpoint (with the block's lifetime, which is 0 when no block matched) and by the swatch media endpoint. The last two are not page views.

2. The built-in full page cache deliberately forbids browser caching. Framework\App\PageCache\Kernel::process() matches /public.*s-maxage=(\d+)/, then calls $response->setNoCacheHeaders() before serialising the response into the cache. So the stored entry carries no-store, no-cache, must-revalidate, max-age=0, and Kernel::load() replays those headers on every hit. A browser is told "never store this" on both a miss and a hit. That is what this module fixes.

3. The built-in cache reads its own storage lifetime out of the s-maxage it finds in that header. Anything that drops public or s-maxage before process() runs silently switches whole-page caching off. This is why the module leaves the declaration moment strictly alone under the built-in cache.

4. Magento's shipped VCL rewrites Cache-Control on the way out — and private is the one value that survives it. module-page-cache/etc/varnish{5,6,7}.vcl, identical in all three:

   That is the very block the VCL snippet above replaces. Its guard is also a door: a response containing private passes through untouched.

What it emits

Pragma becomes cache and Expires is removed in both cases. All three headers move together, because Magento's no-cache set writes all three and any survivor would contradict the other two.

With Varnish as the caching application and a stock VCL, no origin-side solution exists. Anything public is overwritten in vcl_deliver; anything containing private is marked beresp.uncacheable by vcl_backend_response and Varnish stops caching pages altogether. The VCL snippet works only because vcl_deliver runs after Varnish has already stored the object — a moment the origin does not have. The Varnish strategy ships for CDN-fronted installations; if you can edit the VCL, edit the VCL.

The built-in caching application covers both cases this module was written for: no Varnish, and a Varnish you cannot edit.

Features

• Per store view: on/off, lifetime in seconds, must-revalidate, URL exclusions.
• Ships disabled. Installing it changes nothing until an administrator switches it on.
• The browser lifetime is capped at the lifetime Magento announced for the shared cache, so it can never outlive the stored copy.
• Under the built-in cache the response is marked private, so no downstream proxy may keep a copy Magento has no way to purge.
• Magento's own stale-vary guard stays the last word — a page rendered for a context the visitor's browser does not know is never handed a lifetime.
• Five independent eligibility rules, extensible from another module through di.xml without editing a line of this one.

Configuration

Stores → Configuration → Hryvinskyi → Browser Cache Control, ACL resource Hryvinskyi_BrowserCacheControl::config. All fields are shown at default, website and store-view scope and are read at store scope.

Notes that are easy to get wrong:

• browser_ttl of 0 or blank turns the feature off. There is no silent fallback to the default — the config.xml default is the only default, so clearing the field is a way to disable, not a way to reset.
• excluded_url_patterns takes one PCRE body per line, without delimiters. They are compiled with ~ as the delimiter and the i flag, and matched against the request path (leading slash, no query string). A line that does not compile is dropped on read and treated as non-matching — one typo costs one exclusion, never a PHP warning on every page render.
• Magento's own shared lifetime and caching application are global (system/full_page_cache/ttl and …/caching_application are read at default scope), while this module's settings are per store view. The cap reconciles the two.
• The ESI and swatch-media exclusions are correctness invariants, not preferences — they live in etc/di.xml and are deliberately not offered in the admin.

Verifying it

Expect max-age=120, must-revalidate, private under the built-in cache — alphabetically ordered, no s-maxage, no public, no Expires. Run it twice: the first request is a cache miss, the second a hit, and both must carry it. A checkout or customer page must carry none of it.

A signed-in visitor gets nothing on a built-in cache hit, and that is correct rather than a bug: a hit short-circuits before routing, so Framework\App\Http\Context is empty and its vary string is null while the visitor's X-Magento-Vary cookie is a hash. Magento's own HttpPlugin::beforeSendResponse sees the mismatch and strips the caching headers. Guests match null to null and keep them.

How it works

Three plugins, two logical write points, and none of them at response-send time.

Why not beforeSendResponse, the obvious seam: PageCache\Model\App\Response\HttpPlugin::beforeSendResponse strips the caching headers when the visitor's X-Magento-Vary cookie disagrees with the current context. That guard has to keep winning. Writing earlier leaves it the last word for free; writing at send time would mean negotiating plugin sortOrder against core or duplicating core's vary logic. The ordering guarantee comes from which methods are plugged, which is why etc/frontend/di.xml declares no sortOrder at all.

Why the declaration moment for Varnish — nothing has stored the response yet, so what is written there is what the shared cache keeps and obeys. It is the only moment at which a lifetime meant for a shared cache can be announced.

Why not that moment under the built-in cache — see fact 3 above. The built-in path writes only after the copy has been taken, where it changes what the visitor is told and nothing else.

Why Kernel::load() matters most — see fact 2. Without it, the repeat visitor, who has already proved they come back, gains nothing at all.

Layering

The decision and everything below it work on plain values and are unit-tested without Magento present. ApplyBrowserCacheTtl contains no test for the caching application or the write point anywhere — those are asked once, of the strategy. Keeping it that way is the property to protect in review.

The five eligibility rules

All must agree. Order only decides how quickly a rejection is reached.

Developer notes

Adding a caching application

One class implementing Api\PageCacheStrategyInterface — ownsWritePoint(), sharedTtlToPublish(), isPublic() — plus one entry in the strategy pool, keyed by the value stored in system/full_page_cache/caching_application:

No existing file changes. An unrecognised caching application leaves responses alone and logs one warning per request — it never throws, because this runs on every storefront page render.

Adding an exclusion

One class implementing Api\EligibilityRuleInterface — isEligible(ResponseSnapshot): bool — added to the rules argument of Model\ApplyBrowserCacheTtl. Rules never know about each other, so a new reason to skip is a new class rather than a new branch.

Traps worth knowing before you touch this

• xsi:type="number" DI arguments arrive as numeric strings, not integers. Data\Argument\Interpreter\Number::evaluate() returns the raw value uncast, so a strict in_array against int never matches. ResponseStatusRule casts for this reason.
• Laminas\Http\Header\CacheControl::getFieldValue() sorts its directives alphabetically. Whatever CacheDirectives::render() produces, the wire value is max-age=120, must-revalidate, public, s-maxage=86400. Magento's /public.*s-maxage=(\d+)/ still matches only because public happens to sort ahead of s-maxage; core's own header depends on the same accident. Never assert this header against a literal ordered string.
• The ESI full action name is magento_pagecache_block_esi, not page_cache_block_esi. getFullActionName() uses the route id; page_cache is only the front name.
• An announced shared lifetime of 0 is a real announcement, distinct from "unknown". The ESI endpoint announces exactly 0 when no block matched. Reading it as "nothing announced" would fall through to the configured lifetime and hand every unclaimed fragment a lifetime of its own. Every check is !== null, never truthiness.
• PageCache\Model\Config::isEnabled() is the Cache Management toggle, not the caching application, and getTtl() is annotated @return int while returning the raw uncast scope value. Both are normalised once, in Model\PageCacheEnvironment — the only class in this module allowed to name that Magento class.
• On a built-in cache hit the response is not the one AbstractController::getResponse() returns. Kernel::buildResponse() mints a fresh response through HttpFactory, and Framework\App\Http::launch() keeps it in its own private field. Assert on the object Kernel::load() hands back.

Why there is no no-store option

The VCL snippet carries one, commented out, labelled "disable back/forward cache". It is not implemented here and will not be. no-store is unconditional in RFC 9111 §5.2.2.5 — a cache told no-store keeps no part of the response — so pairing it with max-age does not mean "cache briefly but skip the back/forward cache", it means "cache nothing", and the lifetime beside it is dead text. Suppressing the back/forward cache is a consequence of storing nothing, not a capability that can be had while still caching. At the origin it would also be destructive: Magento's vcl_backend_response turns a no-store backend response into hit-for-pass, so Varnish would stop caching the page entirely.

Tests

188 unit tests cover the domain, the rules, both strategies, the decision matrix and all three plugins; 8 integration tests cover the emitted headers under each caching application. The --autoload-file on the phpstan line is required for any Magento integration test: Magento\TestFramework\ is not in composer autoload, and the integration framework self-registers at runtime.

Dependencies

• PHP 8.1 / 8.2 / 8.3 / 8.4 / 8.5
• magento/framework, magento/module-backend, magento/module-config, magento/module-page-cache, magento/module-store

etc/module.xml sequences Magento_PageCache and Magento_Store, so this module's plugins load in the right relationship to Magento's own page cache plugins. Everything is registered in the frontend area only — the admin panel, the web APIs and GraphQL are deliberately untouched.

Author

Volodymyr Hryvinskyi

• Email: [email protected]

License

Proprietary