AEM Sites Playbook, Chapter 11: Delivery, Dispatcher, CDN, domains and caching
Which layer controls the request, redirect or stale response?
November 10, 2026
Draw the route you actually operate
Use this chapter for stale content, unexpected redirects, missing endpoints or differences between an origin and the public site. Record the public domain, TLS termination, redirect rules, CDN, Apache/Dispatcher, origin and any external frontend. Mark ownership of each layer. Edge Delivery or a separately hosted application needs its own route map; do not assume every project uses an identical path.
Record the request method, URL, query parameters, identity and expected content revision. Describe cache policy separately for browser and shared caches. Decide which content may be shared, what varies by user or parameter and how freshness is checked. The Cloud Service caching reference explains its delivery layers; use it to interpret the actual headers and route.
Diagnose one request across layers
Scroll horizontally if needed.
| Observation | Next evidence to collect |
|---|---|
| Redirect before the expected endpoint | Location header, intended method, matching host/rule and where the redirect first appears |
| Missing public content | Intended route, access identity, publication/dependency result and relevant delivery logs |
| Correct origin but stale delivered response | Revision marker, response age/cache headers, route and applicable cache policy |
| User-specific output appears for another identity | Request variation, cache isolation policy and an immediate owner for exposure review |
| Different responses by domain | Virtual-host mapping, TLS/domain configuration, rewrite behavior and the same fixture requested through each domain |
Use approved diagnostics and correlate timestamps/request identifiers without publishing sensitive logs. Change one configuration hypothesis at a time. The POST servlet case reports an .html redirect caused by rewrite handling. Its author-confirmed exclusion is a lesson in route evidence, not a universal allow rule.
Set freshness requirements before choosing a purge
For the hypothetical product update, define when the new fact must be observable and through which endpoint. Compare the content revision at relevant layers. A successful publish, a changed repository property or a cleared browser cache cannot establish the whole result.
The stale persisted-query discussion has no author-confirmed repair, and its accepted purge recipe conflicts with Adobe guidance. Persisted-query documentation allows Dispatcher/CDN caching. For Adobe's managed CDN, the documented purge operation uses PURGE with X-AEM-Purge-Key; URL, key-based and broad purge scopes differ. Follow the current purge reference with authorized credentials and the appropriate scope. A purge is not a fixed end-to-end timing guarantee.
Test the agreed freshness after ordinary publication, withdrawal and relevant cache expiry/revalidation. Include error responses and authorized content in the policy review. Repeated broad flushing can hide a wrong route or cache rule and should not replace that diagnosis.
Convert the symptom into a request comparison
Community discussions about stale responses and servlet routing concern different request paths. The stale-response example has no author-confirmed repair, and its accepted explanation must be checked against current documentation. Compare evidence for the exact operation before generalizing a symptom to other delivery traffic.
Record the request and response at approved diagnostic points. Include method, domain, path, relevant parameters, identity, revision marker, time, status, redirect target and available cache headers. A revision marker can be agreed content evidence; do not add sensitive internal metadata to public output merely for convenience.
Scroll horizontally if needed.
| Finding | Candidate explanation | Evidence that distinguishes it |
|---|---|---|
| The required revision is absent at the expected destination | Publication/dependency problem | Destination/operation and approved content evidence |
| An approved diagnostic response is current but the public response is old | Route/cache variation problem | Same logical request, revision and response headers |
| Browser result differs from a fresh public request | Browser/service-worker/application state may differ | Controlled client conditions and actual response content |
| A POST receives a redirect | Routing may alter the intended operation | Location, actual method flow and matching rule |
| One parameter or identity changes the result | Intended or unintended variation | Cache/access policy and controlled comparison |
Treat each explanation as a hypothesis until the evidence supports it. Similar URLs may select different virtual hosts, consumer logic or cache keys.
Worked scenario of an old product response
All revisions, timings and observations below are hypothetical. A published product fact should appear in a website and a partner consumer by an agreed business deadline. A user reports the website is current but a persisted-query response remains old.
The content owner supplies the approved old/new values. The release owner confirms the relevant content/dependencies and operation result. The delivery and consumer owners agree on equivalent requests and approved diagnostic routes. They do not use unsupported access bypasses.
Scroll horizontally if needed.
| Observation | Illustrative evidence | What it establishes |
|---|---|---|
| Approved content | New value recorded as revision R8 | Intended result |
| Publication | Relevant operation reports completion | One release checkpoint |
| Approved diagnostic response | Required fields reflect R8 | New content is available at that observed point |
| Partner-facing request | Still reflects R7 | Required consumer result has not been met |
| Request comparison | Partner uses an additional parameter and different domain | Requests need route/variation analysis |
The delivery owner traces the actual path for that parameter/domain combination and inspects its policy. If a scoped invalidation is appropriate, the authorized operator follows the documented interface and then repeats the same customer-facing request. No fixed number of seconds is treated as guaranteed proof of freshness.
Suppose the first repair updates the default URL but the partner's parameterized URL remains old. The issue stays open. Testing only the most convenient request would have produced a false success. The team updates the route/test inventory and demonstrates the intended parameterized result before accepting the release.
Choose a cache action with a defined scope
Scroll horizontally if needed.
| Action | Appropriate question | Retest required |
|---|---|---|
| Correct cache policy | Is content varying or expiring differently from intent? | Ordinary hits, variation, revalidation and error behavior |
| Invalidate a known affected object/key | Is the required content already correct at its destination? | Every agreed customer route for the affected scope |
| Correct routing/mapping | Is the request reaching the intended configuration/origin? | Domains, methods, parameters and redirects |
| Correct publication/dependencies | Is the content actually missing before caching? | Publication result and delivered revision |
| Broad invalidation | Is wider scope genuinely necessary and authorized? | Impact, recovery and continued correctness after caches refill |
For Adobe's managed CDN, use the current purge reference. Do not borrow a vendor assumption or a different header/method from an accepted reply. Keep browser, CDN, Dispatcher and consumer/application caches separate in the record.
Include security and failure responses in delivery acceptance
Test authorized and unauthorized identities where relevant, unavailable content, redirects and error responses. Agree which responses may be shared. A performance improvement that exposes one user's content to another is a failed implementation regardless of its speed.
Make the operator's record reproducible: configuration revision, fixture, exact request conditions, response evidence and owner of each layer. After a repair, test an ordinary subsequent publish and withdrawal so correctness does not depend on a one-off manual flush. Update the policy and regression case that should prevent the same incident.
Own the configuration and the result
Use source-controlled configuration and the supported release process for the deployment. Validate relative links, immutable files and tool prerequisites with Dispatcher tools. The local/pipeline discrepancy discussion separates a repaired pipeline issue from an unresolved local one; Windows is not categorically unsupported.
Complete W11 with the route, identity/variation rules, cache policy, redirect tests and observed response. Business owners accept freshness for the customer task; technical owners prove the intended route and recovery. Send metadata/domain consistency to chapter 15 and integration method/authentication requirements to chapter 14.
W11 Request route and cache policy
Use with chapter 11. Compare the same request when investigating different layers.
Scroll horizontally if needed.
| Record | Entry |
|---|---|
| Domain/TLS, frontend/CDN/web-tier/origin route and owners | [enter] |
| Method, URL/parameters, identity and expected revision | [enter] |
| Intended status/redirect and actual Location/header evidence | [enter] |
| Browser/shared cache variation, expiry and invalidation policy | [enter] |
| Freshness/privacy requirement and measured delivered behavior | [enter] |
| Change hypothesis, scoped action, repeat result and rollback | [enter] |
