Deployment and Proxy Placement
In one line: every previous lesson assumed edges exist in the right places. This is where that assumption gets paid for — and placement decisions are what separate Akamai's approach from Google's.
The two questions
- What are the best locations to install proxy servers?
- How many proxy servers are required?
Two placement strategies
| Strategy | Where | Character |
|---|---|---|
| On-premises | Smaller data centers near major IXPs (internet exchange points) | Fewer, larger facilities at network crossroads |
| Off-premises | Proxies placed directly inside ISP networks | Many, smaller footprints — content one hop from users |
Akamai and Google made opposite choices
Akamai and Netflix popularized placing proxies inside client ISPs, often putting content just one hop away. Google relies more on servers near IXPs, likely due to the scale and volatility of its data.
That contrast is the lesson. Inside-ISP placement is unbeatable for latency when your catalogue is large but stable — movies, popular static assets. IXP placement suits data that is volatile or too vast to replicate widely, where you would be constantly invalidating thousands of small ISP footprints.
Content shape drives placement, exactly as it drove push-versus-pull in Lesson 4.
Caching static content such as movies at ISP edge locations is feasible. But services with rapidly growing datasets face a harder question: which content to cache near users at all.
Split TCP
One approach for long-haul delivery is to split TCP.
Client TCP connections terminate at edge infrastructure near IXPs. That edge then forwards traffic to central data centers over persistent, low-latency connections configured with large TCP window sizes.
This reduces client-perceived latency by avoiding the initial TCP three-way handshake and slow-start phases over long distances. Since the round-trip delay to the local IXP is low, the handshake overhead is negligible.
Why split TCP helps even with no caching
This works for content that is not cached at all. TCP slow start means a new connection begins conservatively and ramps up — over a 200 ms path that ramp is brutal, and short transfers finish before reaching full speed.
Splitting gives the client a short, fast-ramping connection to a nearby edge, while the long haul runs over a persistent, already-warmed connection with large windows. You get the benefit of proximity without needing the data to be there.
It's a good example of a CDN improving delivery through transport optimization rather than caching — worth naming, because most candidates only think of CDNs as caches.
Additionally, predictive push strategies help determine which content to proactively move closer to users — pushing based on forecast demand rather than observed requests.
Choosing locations and counts
Placement decisions rely on performance measurements. Tools such as ProxyTeller calculate optimal locations and server counts based on hit ratio, bandwidth, and latency. Other algorithms — greedy, random, and hotspot strategies — are also used.
Why ISPs want the boxes
Placement inside an ISP is a negotiation, and it helps to know the other side's incentives. An ISP benefits because:
| Benefit | Detail |
|---|---|
| Lower transit costs | A CDN node cuts the external bandwidth ISPs need and pay for; SLAs with proxy providers may benefit them economically |
| Happier customers | It improves responsiveness for the ISP's customers, giving a competitive advantage over ISPs without a CDN node |
| Less core traffic | Bringing content closer to users reduces data on the internet core |
Key takeaway
Place edges where connectivity is dense — at IXPs for volatile or vast catalogues, inside ISPs for large but stable ones. And remember that placement improves delivery through transport effects like split TCP, not caching alone.
Interview signal by level
| Level | What a strong answer sounds like |
|---|---|
| L4 | "We put edge servers in different regions." |
| L5 | Knows the two options: "either near IXPs or directly inside ISP networks — inside the ISP puts content about one hop from users." |
| Staff+ | Ties placement to content shape and incentives: "stable catalogues go inside ISPs like Akamai and Netflix; volatile ones sit near IXPs like Google, because you'd otherwise invalidate thousands of footprints constantly. And ISPs take the box because it cuts their transit bill — the incentives align." |
Next: whether to run any of this yourself.