Working for and with organizations in highly regulated industries like federal and state governments and commercial banks often necessitates diving REALLY deep into products and technologies. This means peeling back the layers of the onion most people do not. The reason this pops up is because these organizations tend to have extremely complex environments due the length of time the organization has existed and the strict laws and regulations they must abide by. This is probably the reason why I’ve always gravitated towards these industries.
I recently ran into an interesting use case where that willingness to dive deep was needed.
A customer I was working with was wrapping up its Azure landing zone deployment and was beginning to deploy its initial workloads. A number of these workloads used Microsoft Azure PaaS (platform-as-a-service) services such as Azure Storage and Azure Key Vault. The customer had made the wise choice to consume the services through Azure Private Endpoints. I’m not going to go into detail on the basics of Azure Private Endpoints. There is plenty of official Microsoft documentation that can cover the basics and give you the marketing pitch. You can check out my pasts posts on the topic such as my series on Azure Private DNS and Azure Private Endpoints.
This particular customer chose to use them to consume the services over a private connection from both within Azure and on-premises as well as to mitigate the risk of data exfiltration that exists when egressing the traffic to Internet public endpoints or using Azure Service Endpoints. One of the additional requirements the customer had as to mediate the traffic to Azure Private Endpoints using a security appliance. The security appliance was acting as a firewall to control traffic to the Private Endpoints as well to perform deep packet inspection sometime in the future. This is the requirement that drove me down into the weeds of Private Endpoints and lead to a lot of interesting observations about the behaviors of network traffic flowing to and back from Private Endpoints. Those are the observations I’ll be sharing today.
For this lab, I’ll be using a slightly modified version of my simple hub and spoke lab. I’ve modified and added the following items:
- Virtual machine in hub runs Microsoft Windows DNS and is configured to forward all DNS traffic to Azure DNS (188.8.131.52)
- Virtual machine in spoke is configured to use virtual machine in hub as a DNS server
- Removed the route table from the spoke data subnet
- Azure Private DNS Zone hosting the privatelink.blob.core.windows.net namespace
- Azure Storage Account named mftesting hosting some sample objects in blob storage
- Private Endpoint for the mftesting storage account blob storage placed in the spoke data subnet
The first interesting observation I made was that there was a /32 route for the Private Endpoint. While this is documented, I had never noticed it. In fact most of my peers I ran this by were never aware of it either, largely because the only way you would see it is if you enumerated effective routes for a VM and looked closely for it. Below I’ve included a screenshot of the effective routes on the VM in the spoke Virtual Network where the Private Endpoint was provisioned.
Notice the next hop type of InterfaceEndpoint. I was unable to find the next hop type of InterfaceEndpoint documented in public documentation, but it is indeed related to Private Endpoints. The magic behind that next hop type isn’t something that Microsoft documents publicly.
Now this route is interesting for a few reasons. It doesn’t just propagate to all of the route tables of subnets within the Virtual Network, it also propagates to all of the route tables in directly peered Virtual Networks. In the hub and spoke architecture that is recommended for Microsoft Azure, this means that every Private Endpoint you create in a spoke Virtual Network is propagated to as a system route to route tables of each subnet in the hub Virtual Network. Below you can see a screen of the VM running in the hub Virtual Network.
This can make things complicated if you have a requirement such as the customer I was working with where the customer wants to control network traffic to the Private Endpoint. The only way to do that completely is to create a /32 UDRs (user defined routes) in every route table in both the hub and spoke. With a limit of 400 UDRs per route table, you can quickly see how this may break down at scale.
There is another interesting thing about this route. Recall from effective routes for the spoke VM, that there is a /32 system route for the Private Endpoint. Since this is the most specific route, all traffic should be routed directly to the Private Endpoint right? Let’s check that out. Here I ran a port scan against the Private Endpoint using nmap using the ICMP, UDP, and TCP protocols. I then opened the Log Analytics Workspace and ran a query across the Azure Firewall logs for any traffic to the Private Endpoint from the VM and lo and behold, there is the ICMP and UDP traffic nmap generated.
Yes folks that /32 route is protocol aware and will only apply to TCP traffic. UDP and ICMP traffic will not be affected. Software defined networking is grand isn’t it? 🙂
You may be asking why the hell I decided to test this particular piece. The reason I followed this breadcrumb was my customer had setup a UDR to route traffic from the VM to an NVA in the hub and attempted to send an ICMP Ping to the Private Endpoint. In reviewing their firewall logs they saw only the ICMP traffic. This finding was what drove me to test all three protocols and make the observation that the route only affects TCP traffic.
Microsoft’s public documentation mentions that Private Endpoints only support TCP at this time, but the documentation does not specify that this system route does not apply to UDP and ICMP traffic. This can result in confusion such as it did for this customer.
So how did we resolve this for my customer? Well in a very odd coincidence, a wonderful person over at Microsoft recently published some patterns on how to approach this problem. You can (and should) read the documentation for the full details, but I’ll cover some of the highlights.
There are four patterns that are offered up. Scenario 3 is not applicable for any enterprise customer given that those customers will be using a hub and spoke pattern. Scenario 1 may work but in my opinion is going to architect you into a corner over the long term so I would avoid it if it were me. That leaves us with Scenario 2 and Scenario 4.
Scenario 2 is one I want to touch on first. Now if you have a significant background in networking, this scenario will leave you scratching your head a bit.
Notice how a UDR is applied to the subnet with the VM which will route traffic to Azure Firewall however, there is no corresponding UDR applied to the Private Endpoint. Now this makes sense since the Private Endpoint would ignore the UDR anyway since they don’t support UDRs at this time. Now you old networking geeks probably see the problem here. If the packet from the VM has to travel from A (the VM) to B (stateful firewall) to C (the Private Endpoint) the stateful firewall will make a note of that connection in its cache and be expecting packets coming back from the Private Endpoint representing the return traffic. The problem here is the Private Endpoint doesn’t know that it needs to take the C (Private Endpoint) to B (stateful firewall) to A (VM) because it isn’t aware of that route and you’d have an asymmetric routing situation.
If you’re like me, you’d assume you’d need to SNAT in this scenario. Oddly enough, due the magic of software defined routing, you do not. This struck me as very odd because in scenario 3 where everything is in the same Virtual Network you do need to SNAT. I’m not sure why this is, but sometimes accepting magic is part of living in a software defined world.
Finally, we come to scenario 4. This is a common scenario for most customers because who doesn’t want to access Azure PaaS services over an ExpressRoute connection vs an Internet connection? For this scenario, you again need to SNAT. So honestly, I’d just SNAT for both scenario 2 and 4 to make maintain consistency. I have successfully tested scenario 2 with SNAT so it does indeed work as you expect it would.
Well folks I hope you found this information helpful. While much of it is mentioned in public documentation, it lacks the depth that those of us working in complex environments need and those of us who like to geek out a bit want.
See you next post!