Scaling AEM Dispatcher Farms with Elastic IP on AWS

Adobe Experience Manager runs the digital estates of plenty of Australian organisations — a bank's internet banking app, an insurer's claims portal, a telco's self-care journey — and almost all of that content sits behind Apache Dispatcher rather than the authoring tier. When traffic grows, the instinct is to add more dispatcher instances, and on Amazon Web Services that decision has a surprisingly physical consequence: every new instance arrives with a new IP address. Elastic IP addresses soften that shock, and understanding where they belong in a dispatcher farm is the difference between a scale-out that takes minutes and one that stalls behind firewall rules and DNS caches.

The trouble usually starts at the publish tier. Dispatchers cache rendered content, invalidate it through a flush agent, and forward cache misses to authoring instances sitting behind a load balancer in a private subnet. Those publishers call third-party systems — a DAM, a product API, a marketing automation platform — and those integrations are frequently pinned to a short list of permitted addresses. Change the address of your origin and an outage follows. In Australia that risk is amplified by latency expectations, data residency requirements, and hybrid connectivity into on-premises systems that nobody wants to re-address.

An Elastic IP is a static, routable address allocated from Amazon's pool. You associate it with an instance, disassociate it, and associate it with a different instance within seconds, so a hostname can keep resolving to the same address while the workload behind it rotates. That makes it a useful anchor for a dispatcher farm, provided you are clear about the parts it does not solve. Elastic IPs do not cache content, do not balance traffic, and do not health check. They are addressing glue. Autoscaling, cache consistency, session affinity and cost control still have to be designed — and those are exactly the subjects the speaker lineup at CIRCUIT has wrestled with in the past.

Why dispatcher farms strain under real traffic

Australian traffic is not flat. Sydney and Melbourne offices generate a sharp 7am to 9am peak as commuters tap on with an Opal card or a Myki pass and open the morning headlines, while Brisbane, Perth and Adelaide follow a slightly later curve. Add a public holiday on 25 January or 25 April, a cricket Test at the MCG, or the 30 June end-of-financial-year rush, and peaks get taller and far less predictable. A farm sized for average load will thrash through those windows.

NBN and mobile connectivity have shifted expectations too. Australians now expect pages that render in a blink on a phone tethering in a lift, which means cache hit rates matter more than raw CPU. Every miss at the dispatcher layer is a request forwarded to a publisher that also has to call an external API. Scaling the farm without scaling the cache tier simply moves the bottleneck downstream, where you have less visibility and less control.

The other strain is operational. When dispatchers scale out, the flush agent — usually a reverse proxy sitting in front of the farm — must invalidate every node. A node that registers late will serve stale content for a window that can last hours. Teams that have lived through that learn to treat cache consistency as the first design question rather than the last, and to keep the flush agent configuration in version control instead of on one person's server.

Elastic IP as a stable front door for a moving fleet

An Elastic IP stays with the account, not the machine. You associate one with each dispatcher, or hand the address to a network load balancer target group, and rebuild the origin configuration whenever you want to replace the instance behind it. That matters for two specifically Australian reasons: content delivery providers and security vendors who whitelist origin addresses, and hybrid links — site-to-site VPNs or Direct Connect circuits into a corporate datacentre — where routes to a fixed prefix are documented, audited and occasionally audited again by a client.

A load balancer target group can register Elastic IP addresses as targets where the address belongs to an instance in the same VPC, and a health check against the dispatcher's /cache.status endpoint can pull a failing node out of rotation before it serves errors. If your architecture mixes on-premises dispatchers with cloud ones, you can register those addresses too, provided the load balancer can route to them. Dennis Ideler's write-up walks through the remapping mechanics and the DNS caching gotcha that follows: a hostname with a five-minute TTL will keep sending traffic to a decommissioned instance for five minutes after the address moves.

Two constraints deserve attention. Elastic IP quotas are small by default — five per region in most accounts — and raises require a support request, so request headroom before the conference season rather than during it. An address that sits unassociated also stops being free, which turns forgotten IPs into a small, puzzling line on the AWS bill. Separately, government and regulated clients in Australia increasingly ask about IRAP-assessed cloud services and the notifiable data breaches regime under the Privacy Act 1988, which means an origin that changes address without notice is both an availability risk and a reporting risk.

Autoscaling publishers without breaking sessions

Publisher instances belong in an autoscaling group, scaled on request count or CPU, with a health check that reports unhealthy while an instance is still warming its dispatcher cache. Warm-up matters more than it sounds: a dispatcher that has not synchronised its cache forwards a flood of misses to the publisher, and the scaling event becomes self-reinforcing. Scale on the number of requests the publisher is actually serving, and give new nodes a grace period before the flush agent starts invalidating through them.

Authoring traffic behaves differently. Granite UI screens, AEM Forms sessions and AEM Assets workflows keep state on the instance, so author nodes need affinity and a stable address. This is where Elastic IP combined with a target group earns its keep: a session can be pinned to an address that survives the underlying instance being replaced, and the address a partner agency whitelisted last quarter is still the address it whitelists next quarter. Long-lived connections need their own treatment, since every remap breaks them.

The underlying maths is not new. The open-source scalability work published in the ISORC 2015 papers still frames the fundamentals that matter here: measured saturation points, cost per useful request rather than cost per instance, and how quickly a fleet degrades when every node competes for the same downstream API with a per-second quota.

Practical recommendations for Australian AEM teams

Before committing to an addressing model, work through the following.

  • Keep one Elastic IP per dispatcher origin and re-associate it rather than re-pointing DNS, timing changes around your CDN's TTL and purge hooks.
  • Put /cache.status behind a load balancer health check so broken nodes leave rotation before they serve stale pages or 500s.
  • Raise your Elastic IP quota for the region up front; the default allocation is too small for a farm with separate author and publish tiers.
  • Document the addresses in hybrid routing tables and third-party allow lists, then rehearse a full failover before December, before EOFY, and before the Sydney Marathon morning.
  • Version the flush agent configuration and confirm it reaches every dispatcher, including nodes added five minutes ago.

Bring your dispatcher farm to the next conference

Most of this comes from people who run these estates every day, and CIRCUIT was built for exactly that conversation. The Chicago programme covered AEM integrations, Sightly templates, mobile delivery, microservices and architecture, with sessions on scaling publisher and author tiers in the cloud and on building a dispatcher farm that survives a peak.

If you want the practical details — dates, venue, ticket pricing history and recordings from the 2015 and 2016 editions — the conference FAQ covers the common questions and points you at the agenda and speaker pages for each track.

Bring three things: a diagram of your current dispatcher topology, a copy of your flush agent configuration and your AWS quota request. That is enough to turn a hallway conversation with architects, engineers and front-end developers into a working plan for the next growth spike. Register for CIRCUIT, or share your own scaling notes with the community and keep the conversation going after the lights go down.