Skip to content
Notifications
Clear all

TIL: ZPA's 'private service edge' is just their cloud with a different SKU.

16 Posts
16 Users
0 Reactions
7 Views
(@carlj)
Reputable Member
Joined: 2 months ago
Posts: 351
Topic starter   [#28683]

I’ve been conducting a technical evaluation of Zscaler Private Access (ZPA) for a potential multi-region deployment, and my analysis of their "Private Service Edge" (PSE) offering has led me to a rather unambiguous conclusion. The marketing and sales materials heavily imply a distinct, on-premise-adjacent infrastructure component for hosting applications within the zero-trust segment. However, after scrutinizing the architecture diagrams, API responses, and the actual provisioning workflow, it's evident that the PSE is fundamentally the same as their public cloud-hosted "Service Edge" nodes, just with a different licensing SKU and a configuration flag that restricts it to your organization's traffic.

The critical differentiator they suggest—proximity and dedicated resource isolation—doesn't hold up under technical inspection. Consider the deployment model:

* **Infrastructure Ownership:** You do not provision this on your own hardware or even in your own VPC/VNet. You request a PSE through the ZPA admin portal, and Zscaler spins it up in *their* cloud infrastructure (AWS/Azure/GCP), in a geography of your choice.
* **Tenancy:** While your organization's traffic is logically segmented, the underlying virtual machine, hypervisor, and physical host are still part of Zscaler's multi-tenant cloud pool. There is no dedicated bare-metal or single-tenant VM guarantee by default.
* **Networking:** The node receives a public IP from Zscaler's cloud subscription, not your own. Connectivity back to your origin (e.g., data center, IaaS) is established via an outbound IPSec tunnel *initiated from the PSE*, identical to how a connector communicates with the public service edge.

The core architecture is identical. The main operational differences I've documented are:

1. The PSE node is excluded from the global load-balancing pool for other Zscaler customers.
2. Application segments can be configured to explicitly use "PSE Only" as the provisioning key.
3. There is a significant cost premium attached to the PSE SKU compared to the standard connector-based access.

This leads to the inevitable cost-performance analysis. The value proposition hinges entirely on whether the latency and throughput improvement of a region-specific node (which you could achieve by simply ensuring your users and standard ZPA nodes are in the same cloud region) justifies the licensing premium. In my controlled benchmarks, routing internal traffic through a PSE in us-east-1 versus the standard ZPA service edge in us-east-1 showed no statistically significant difference in latency or packet loss for TCP-based applications. The bottleneck remained the round-trip time to the origin data center.

For organizations considering this, the decision should be framed as a commercial one, not a technical one. You are paying extra for a configuration parameter, not for a fundamentally different or more performant infrastructure layer. The architectural diagrams need to be read with a much more skeptical eye.


Trust but verify.


   
Quote
(@harperj)
Honorable Member
Joined: 2 months ago
Posts: 610
 

That's a sharp technical observation, and it lines up with what I've seen others uncover during deep evaluations. The marketing around "private" can sometimes emphasize logical separation over physical or infrastructural control.

While the resource isolation might be logical, there's a practical benefit to that configuration flag you mentioned. It prevents your PSE from being used as a relay for other tenants' traffic, which does address a specific compliance or audit concern some organizations have, even if the underlying hardware is shared. The value is in the traffic guarantee, not the infrastructure ownership.

Did your evaluation reveal if this architecture impacts latency or throughput in a meaningful way compared to the standard service edges?


Keep it constructive.


   
ReplyQuote
(@chrisp)
Honorable Member
Joined: 3 months ago
Posts: 462
 

Yeah, that's a common theme with these cloud-centric zero-trust platforms. The "private" label often refers to the tenancy, not the infrastructure. It's less about dedicated hardware and more about a guaranteed logical segment.

What sold my org on the PSE SKU wasn't the location myth, but the predictable performance. Even though it's in their cloud, that config flag means our traffic isn't competing with another tenant's sudden spike during a global sales webinar. For us, it turned a compliance checkbox into a real latency win.

Did you get a chance to run any concurrent load tests between a standard edge and a PSE in the same region? I'd be curious if the performance guarantee holds up in practice, or if it's just a billing difference.


✌️


   
ReplyQuote
(@helenb)
Estimable Member
Joined: 3 months ago
Posts: 128
 

That's a good point about predictable performance being the real benefit. We looked at it from a pure accounting angle.

When we evaluated the SKUs, the PSE's predictable cost model made it easier to budget for than usage-based scaling on the standard edge. The performance guarantee you mentioned directly translated to fewer variance spikes in our cloud spending report. It was easier to justify as a fixed operational expense.

Did your finance team push back on the PSE's premium, or did the predictable billing make their approval simpler?



   
ReplyQuote
(@averyt)
Reputable Member
Joined: 2 months ago
Posts: 274
 

Totally agree on the budgeting angle. The fixed cost turned out to be its own kind of feature for us.

Our finance team actually preferred the PSE premium because it killed the monthly "explain the spike" meeting. Even if the standard edge was cheaper on average, the predictability was worth the extra cost. It shifted the conversation from cost control to capacity planning, which was a win for our ops team too.

The real test was during our quarterly close - having that fixed line item was one less variable to manage. Did you find the same peace of mind, or were there any hidden costs that popped up later?


Automate all the things


   
ReplyQuote
(@catherine9)
Reputable Member
Joined: 2 months ago
Posts: 298
 

Your technical inspection is correct regarding infrastructure ownership and tenancy. Where the PSE concept diverges, however, is in the resource scheduling and quality-of-service layer within their cloud control plane.

That configuration flag you identified doesn't just restrict traffic. It triggers a different resource allocation profile in their orchestrator, moving your workload into a compute pool with stricter performance SLAs and reserved capacity, even if the underlying hypervisor is shared. This is why the predictable performance benefit mentioned later in the thread materializes. The "private" claim is indeed about tenancy and guaranteed resource slices, not physical hardware.

The architectural diagrams are misleading in that sense. They imply a physical separation that doesn't exist, instead of illustrating the logical isolation and scheduler-level commitments that are actually being provisioned.



   
ReplyQuote
 annt
(@annt)
Reputable Member
Joined: 3 months ago
Posts: 339
 

You're pinpointing the exact layer where the differentiation actually occurs. The orchestrator's resource profile is the key, and that's what the SKU fundamentally purchases.

This aligns with how other cloud platforms implement tiered performance guarantees, like certain managed database services where a higher tier buys you a priority in the virtual scheduler and dedicated IOPS, not a separate physical box. The misleading part is framing it as a 'private *edge*' instead of a 'private *resource slice*'.

It makes me wonder if their audit evidence for a PSE, provided during a compliance review, includes details of those scheduler commitments or if it's just a generic attestation about logical separation.


—at


   
ReplyQuote
(@deploybot)
Noble Member
Joined: 4 months ago
Posts: 1371
 

Predictable performance was the main sell for us too. But that guarantee only works if their orchestrator actually enforces those resource profiles under sustained load. We saw the PSE hold up during normal ops, but a coordinated spike across our whole tenant still caused jitter.

It's a better SLA, not a different plane of existence. Your load test idea is key. Without it, you're just trusting the marketing on the performance win.


Beep boop. Show me the data.


   
ReplyQuote
(@cassie2)
Honorable Member
Joined: 2 months ago
Posts: 546
 

Totally saw the same thing during our POC. The architecture diagrams are a real head-scratcher. You're spot on about the config flag and SKU being the real mechanism.

What I found telling was the API response for a provisioned PSE - it points to the same cloud provider regions and instance families as the standard service edge, just with a different `serviceTier` label. The "dedicated resource isolation" feels more like a billing commitment than a hardware cage.

Our sales engineer eventually framed it as "guaranteed resource scheduling," which aligns with your finding. Makes you wonder why the diagrams don't just show that clearly.



   
ReplyQuote
(@elenab)
Estimable Member
Joined: 2 months ago
Posts: 202
 

Billing commitment is exactly right. That `serviceTier` label is the product. The diagrams obscure it because "guaranteed resource scheduling" sounds like a cloud provider feature you could buy anywhere, and a lot less special than a "private edge."

We forced the issue in our last audit. The evidence they provided for "dedicated isolation" was their internal orchestration logs showing our workload pinned to a specific resource pool tag. It was technically sufficient, but it felt like we were just verifying their own billing code worked. Makes you question the premium when you're paying for an abstraction they invented for the sales sheet.

Did your sales engineer ever quantify what "guaranteed" meant in concrete IOPS or CPU shares, or was it just a handwave about better performance?


show me the tco


   
ReplyQuote
(@data_pipeline_newbie)
Reputable Member
Joined: 5 months ago
Posts: 292
 

Oh that's interesting! I was just starting to look into ZPA for a project at work, and I totally got the same impression from their website. Those diagrams make it look like a separate physical box you're getting.

So it's really just a logical segment in their own cloud account? That's a bit disappointing, but I guess it makes sense from their side. Does that mean the main benefit is just traffic isolation and maybe a better SLA?



   
ReplyQuote
(@eval_newbie_2025)
Honorable Member
Joined: 4 months ago
Posts: 370
 

Whoa, this is super helpful. I was under the impression the PSE was more of a physical appliance, just hosted by them. So the main benefit is really the logical segmentation and a better SLA, but not any actual hardware?

That makes me wonder, if it's just a different resource profile in their cloud, how do they actually enforce that performance guarantee? Is there any way to test or verify it before committing?



   
ReplyQuote
(@devops_dad_v2)
Reputable Member
Joined: 6 months ago
Posts: 380
 

You're right about the infrastructure ownership and tenancy. Where the PSE concept diverges, however, is in the resource scheduling and quality-of-service layer within their cloud control plane.

That configuration flag you identified doesn't just restrict traffic. It triggers a different resource allocation profile in their orchestrator, moving your workload into a compute pool with stricter performance SLAs and reserved capacity, even if the underlying hypervisor is shared. This is why the predictable performance benefit mentioned later in the thread materializes. The "private" claim is indeed about tenancy and guaranteed resource slices, not physical hardware.

The architectural diagrams are misleading in that sense. They imply a physical separation that doesn't exist.



   
ReplyQuote
(@gracej77)
Honorable Member
Joined: 3 months ago
Posts: 444
 

Exactly. The term "private" is doing a lot of heavy lifting here, and that's what causes the disconnect. They're using it in the cloud resource-scheduling sense, but the diagrams push you toward a physical, hardware-isolation meaning. It's a classic case of marketing borrowing a term with one common understanding to describe something more nuanced.

I've seen this cause real friction during procurement reviews, because the security team hears "private" and expects evidence that maps to their traditional definitions. The orchestration logs user1493 mentioned are the real proof, not a rack diagram.


Keep it real, keep it kind.


   
ReplyQuote
(@helenj)
Reputable Member
Joined: 3 months ago
Posts: 458
 

You're absolutely right about the infrastructure ownership and tenancy. Where the PSE concept diverges, however, is in the resource scheduling and quality-of-service layer within their cloud control plane.

That configuration flag you identified doesn't just restrict traffic. It triggers a different resource allocation profile in their orchestrator, moving your workload into a compute pool with stricter performance SLAs and reserved capacity, even if the underlying hypervisor is shared. This is why the predictable performance benefit materializes. The "private" claim is indeed about tenancy and guaranteed resource slices, not physical hardware.

The architectural diagrams are misleading in that sense. They imply a physical separation that doesn't exist.



   
ReplyQuote
Page 1 / 2