Appearance
Cloud Networking
530. What is a VPC or VNet?
The short answer
A VPC, virtual private cloud, in AWS and Google Cloud, or a VNet, virtual network, in Azure, is your own private, isolated network inside the cloud. You choose its address range, divide it into subnets, control its routes and firewall rules, and decide what connects to the internet or to your offices. Resources such as virtual machines, databases and Kubernetes nodes live inside it, and by default nothing outside can reach them.
Think of it as renting an empty office floor in a huge building. The building, the cables and the power belong to the provider. Inside your floor, you decide where the walls go, who gets a key, and which doors open to the street.
What you control inside it:
- The address range, such as
10.20.0.0/16(question 531). - Subnets: smaller ranges for different tiers, such as web, app and database (question 532).
- Routes: where traffic for each destination goes (question 533).
- Firewall rules: security groups in AWS, network security groups in Azure (question 534).
- Connections: to the internet, to other networks, and to your data centre (questions 535 and 536).
One important difference between clouds:
| AWS VPC | Azure VNet | Google VPC | |
|---|---|---|---|
| Spans | One region | One region | Global |
| Subnets span | One availability zone each | The whole region | One region each |
So in AWS, spreading across three zones means at least three subnets per tier. In Azure, one subnet can hold machines in all zones.
Isolation by default. Two VPCs, even in the same account, cannot talk to each other until you connect them. That is the basis of keeping environments and customers apart.
531. How do CIDR blocks and subnet sizing work in cloud networks?
The short answer
CIDR notation writes an address range as a starting address and a prefix length, such as 10.20.0.0/16. The number after the slash says how many of the 32 bits are fixed, so the rest are free for addresses: /16 leaves 16 bits, about 65,000 addresses, and /24 leaves 8 bits, 256 addresses. Choose ranges that do not overlap with any network you may ever connect to, leave room to grow, and remember that clouds reserve a few addresses in every subnet.
How to read a CIDR block. An IPv4 address is 32 bits, written as four numbers from 0 to 255. The prefix length says how many bits, from the left, are the network part. The remaining bits number the hosts.
| CIDR | Free bits | Addresses | Example use |
|---|---|---|---|
/16 | 16 | 65,536 | A whole VPC or VNet |
/20 | 12 | 4,096 | A big subnet, such as Kubernetes nodes and pods |
/24 | 8 | 256 | A typical subnet |
/27 | 5 | 32 | A small subnet, such as for a gateway |
/32 | 0 | 1 | One single address, in firewall rules |
A worked example. VNet 10.20.0.0/16 split into subnets:
text
10.20.0.0/24 web tier 10.20.0.0 – 10.20.0.255
10.20.1.0/24 app tier 10.20.1.0 – 10.20.1.255
10.20.2.0/24 database tier 10.20.2.0 – 10.20.2.255
10.20.16.0/20 Kubernetes nodes 10.20.16.0 – 10.20.31.255Reserved addresses. Clouds keep some addresses in every subnet for their own use: Azure and AWS both reserve five per subnet, such as the network address and the gateway. A /24 gives you 251 usable addresses, and a /29, 8 addresses, gives only 3.
Sizing rules that save pain later:
- Never overlap with your offices, other VPCs, partner networks or VPN ranges. Overlapping networks cannot be connected directly, and changing a range later usually means rebuilding the network.
- Plan big for Kubernetes. When pods get addresses from the network, as with Azure CNI's pod-subnet mode or AWS VPC CNI, every pod uses one. A small subnet runs out fast (question 614).
- Leave space for subnets you have not thought of yet: private endpoints, gateways, future services.
- Use private ranges:
10.0.0.0/8,172.16.0.0/12and192.168.0.0/16are reserved for private networks and are never routed on the internet.
532. What is the difference between public and private subnets?
The short answer
A public subnet's resources can be reached from, and can reach, the internet directly, because the subnet has a route to an internet gateway and resources there can have public addresses. A private subnet has no direct route in from the internet. Its resources keep private addresses only, and if they need to reach the internet, for example to download updates, they go out through a NAT gateway, which allows outbound connections but no inbound ones. Put only entry points, such as load balancers, in public subnets.
| Public subnet | Private subnet | |
|---|---|---|
| Route to the internet | Direct, through an internet gateway | None, or outbound only through NAT |
| Public IP addresses | Possible | No |
| Reachable from the internet | Yes, if firewall rules allow | No |
| Put here | Load balancers, NAT gateways, bastions | App servers, databases, Kubernetes nodes |
What actually makes a subnet "public" in AWS is its route table: a route sending 0.0.0.0/0, meaning all internet traffic, to an internet gateway. In Azure, subnets have default outbound internet access unless you change it, and new guidance is to make outbound access explicit through a NAT gateway, while inbound reachability is controlled by public IP addresses and network security groups.
The principle: the fewer things with public addresses, the fewer things an attacker can reach directly. Ideally, only the load balancer, or a gateway service in front of it, faces the internet.
533. What are internet gateways, NAT gateways, and route tables?
The short answer
A route table is a list of rules saying where traffic for each destination address range should go next. An internet gateway is the door between a VPC and the internet, used by resources with public addresses, in both directions. A NAT gateway lets resources with only private addresses start connections to the internet, by replacing their private source address with the gateway's own public address, while never letting the internet start a connection back in.
A route table, for an AWS private subnet:
| Destination | Target | Meaning |
|---|---|---|
10.20.0.0/16 | local | Anything inside the VPC: deliver directly |
10.30.0.0/16 | Peering connection | The partner VPC: through the peering |
0.0.0.0/0 | NAT gateway | Everything else, the internet: through NAT |
The most specific matching route wins: traffic for 10.30.4.7 matches 10.30.0.0/16 before 0.0.0.0/0.
How NAT works, step by step:
- An app server at
10.20.1.15connects toapi.stripe.com. - The route table sends the packet to the NAT gateway.
- The NAT gateway rewrites the source address to its own public address, say
20.61.8.9, remembers the connection, and sends it on. - The reply comes back to
20.61.8.9. The NAT gateway looks up the connection and forwards the reply to10.20.1.15. - A packet arriving at
20.61.8.9that belongs to no connection the inside started is dropped. That is why NAT allows out but not in.
Practical points:
- One stable outbound address. Partners who only accept calls from known addresses can allow-list the NAT gateway's public IP.
- Cost. NAT gateways charge per hour and per gigabyte processed. Heavy traffic to cloud services, such as pulling images from the registry, can be routed privately instead, through private endpoints, to save money (question 538).
- Availability. In AWS, a NAT gateway lives in one zone: for resilience, use one per zone.
534. What is the difference between security groups and network ACLs or network security groups?
The short answer
They are both firewall rules, at different points. An AWS security group is attached to a resource, such as a virtual machine's network interface, is stateful, so replies to allowed traffic are allowed automatically, and only contains allow rules. An AWS network ACL is attached to a subnet, is stateless, so return traffic needs its own rule, and has numbered allow and deny rules. Azure's network security group is stateful, with prioritised allow and deny rules, and can be attached to a subnet, a network interface, or both.
| AWS security group | AWS network ACL | Azure network security group | |
|---|---|---|---|
| Attached to | Network interface | Subnet | Subnet and/or network interface |
| Stateful | Yes | No | Yes |
| Rule types | Allow only | Allow and deny | Allow and deny |
| Rule order | All evaluated | Lowest number first, first match wins | Lowest priority number first |
| Can reference other groups | Yes, "from the web tier's group" | No, addresses only | Yes, through application security groups |
Stateful versus stateless, in plain words. A stateful firewall remembers connections. If it allowed an outgoing request, it lets the reply in, without a separate rule. A stateless firewall judges every packet on its own, so you must allow the replies explicitly, including the high-numbered temporary ports that replies use.
An example, security groups referencing each other:
text
sg-web: allow inbound 443 from 0.0.0.0/0
sg-app: allow inbound 8080 from sg-web
sg-db: allow inbound 5432 from sg-appThe database accepts connections only from machines in the app tier's group, whatever their addresses. New app servers join sg-app and are allowed automatically.
In practice: security groups, or network security groups, do almost all the work, because they are stateful and precise. Network ACLs are a coarse extra layer, for example to block a range of addresses for a whole subnet.
535. What is the difference between VPC peering and a transit gateway or hub-and-spoke network?
The short answer
Peering connects exactly two networks directly, so their resources can talk as if on one network. It is simple and fast, but it is not transitive: if A peers with B and B with C, A still cannot reach C, so many networks need many peerings. A transit gateway, or a hub-and-spoke design, puts one central hub in the middle that every network connects to, so routing, shared services, firewalls and the link to your offices are managed in one place.
| Peering | Transit gateway, or hub and spoke | |
|---|---|---|
| Connects | Two networks | Many networks, through a centre |
| Transitive | No | Yes, through the hub's routing |
| Links needed for n networks | Up to n × (n − 1) / 2 | n |
| Central firewall and inspection | No | Yes |
| Shared office connection | Each network separately | Once, at the hub |
| Cost | Mostly data transfer | Hub hourly charges plus data |
The products: AWS has VPC peering and Transit Gateway. Azure has VNet peering, and builds hub and spoke either by hand, with a hub VNet containing Azure Firewall and the VPN gateway, peered to each spoke, or with Virtual WAN, a managed hub. Google Cloud has VPC peering, Shared VPC, and Network Connectivity Center.
Overlapping ranges cannot be peered, which is one more reason to plan address ranges centrally (question 531).
536. What is the difference between a site-to-site VPN and a dedicated connection like Direct Connect or ExpressRoute?
The short answer
Both connect your offices or data centre to your cloud network. A site-to-site VPN sends encrypted traffic over the ordinary internet: it is quick to set up and cheap, but its speed and delay vary with the internet. A dedicated connection, AWS Direct Connect or Azure ExpressRoute, is a private physical link through a connectivity provider, and it never crosses the public internet. It costs more and takes weeks to arrange. In return it gives steady bandwidth and low, predictable delay, and it is often required for large or sensitive traffic.
| Site-to-site VPN | Direct Connect, ExpressRoute | |
|---|---|---|
| Path | Encrypted tunnel over the internet | Private circuit through a provider |
| Set-up time | Hours | Weeks, needs a physical connection |
| Bandwidth | Up to a few Gbps, variable | 1 to 100 Gbps, consistent |
| Latency | Varies with the internet | Low and predictable |
| Encryption | Built in, IPsec | Not by default; add it if required |
| Cost | Low | Higher: port and circuit fees |
| Common use | Small sites, backup link, quick start | Large data transfer, core business traffic |
IPsec is the standard protocol suite VPNs use to encrypt traffic between two gateways.
A common pattern: a dedicated connection as the main link, with a VPN over the internet as a backup, so a cut circuit does not cut the company off from the cloud.
Routing between them usually uses BGP, the Border Gateway Protocol, which is the routing protocol of the internet. With BGP, each side learns automatically which address ranges live on the other side. If the main link fails, traffic moves to the backup link by itself.
537. How can resources in a private network reach cloud services such as storage without going over the internet?
The short answer
Clouds offer two ways. Service endpoints, or gateway endpoints in AWS, keep traffic from your subnet to a cloud service on the provider's own network and let the service restrict access to your subnet, while the service keeps its public address. Private endpoints go further: the service gets a private address inside your own network, so you reach it exactly like any other internal resource, and its public access can be turned off completely.
| Service endpoint or gateway endpoint | Private endpoint, PrivateLink | |
|---|---|---|
| Service address | Still the public one | A private IP in your subnet |
| Traffic path | The provider's backbone | Your network to the service privately |
| Reachable from your offices over VPN | Usually not | Yes |
| Public access can be disabled | Partly, by subnet rules | Yes, completely |
| DNS changes needed | No | Yes (question 538) |
| Cost | Usually free | Per endpoint per hour, and per gigabyte |
Examples:
- AWS gateway endpoints, for S3 and DynamoDB, add a route to the service through the VPC's route table. They are free, and save NAT gateway charges for heavy S3 traffic.
- Azure service endpoints let a storage account accept traffic only from chosen subnets.
- Private endpoints, called PrivateLink in AWS and Azure Private Link in Azure, put a network interface for the service into your subnet.
Why it matters: sensitive data, such as backups or customer files, never needs to be reachable from the internet at all. With private endpoints, a storage account or database can refuse every public connection, which removes a whole category of mistakes, such as an accidentally public bucket (question 557).
538. How do private endpoints work, and why does DNS matter for them?
The short answer
A private endpoint is a network interface in your subnet, with a private IP address, that leads privately to one specific cloud resource, such as your registry or database. Clients still use the resource's normal name, such as myacr.azurecr.io, so DNS must answer that name with the private address for clients inside your network. That is done with a private DNS zone linked to your networks. If DNS still returns the public address, traffic goes to the public endpoint, which may be blocked, and connections fail.
myacr.azurecr.io. ② The private DNS zone answers with the private endpoint's address. ③ The node connects there. ④ Traffic reaches the registry without touching the internet.How the name resolves. When a private endpoint is created for myacr.azurecr.io, public DNS starts answering that name with an alias, a CNAME, to myacr.privatelink.azurecr.io. Then:
- Inside your network, a private DNS zone called
privatelink.azurecr.io, linked to your VNet, answers that alias with the private address10.20.5.4. - Outside, public DNS answers with the public address, which refuses connections if public access is disabled.
What breaks, and why:
- The private DNS zone is not linked to the VNet where the client is. The client gets the public address, and the connection is refused or times out.
- Clients use custom DNS servers, such as company DNS servers, which do not know about the private zone. They need to forward these names to the cloud's resolver.
- Clients in your offices, over VPN or ExpressRoute, also need DNS forwarding to resolve private names.
- Hub and spoke: the private zone must be linked to every network whose clients use it.
A quick check from a client:
bash
nslookup myacr.azurecr.ioThe answer should be a private address such as 10.20.5.4. A public address means the DNS set-up is the problem, not the network (question 160).
539. What is the difference between layer 4 and layer 7 load balancers in the cloud?
The short answer
A layer 4 load balancer works with connections: it looks at addresses and ports, and spreads TCP or UDP connections across backends, without reading what is inside. It is very fast and works for any protocol. A layer 7 load balancer works with requests: it understands HTTP, so it can end TLS, route by host name and URL path, add headers, and apply a web application firewall. Use layer 4 for raw speed and non-HTTP traffic, and layer 7 for web applications and APIs.
The "layers" come from the OSI model, a standard way of describing networks in seven layers. Layer 4 is the transport layer, TCP and UDP. Layer 7 is the application layer, such as HTTP.
| Layer 4 | Layer 7 | |
|---|---|---|
| Understands | Addresses, ports, connections | HTTP requests: hosts, paths, headers |
| Protocols | Any TCP or UDP | HTTP, HTTPS, gRPC, WebSockets |
| TLS | Usually passed through | Usually ended here |
| Routing | By port | By host name, path, header |
| Extras | Very high throughput, static addresses | WAF, redirects, rewrites, cookies |
| AWS | Network Load Balancer | Application Load Balancer |
| Azure | Load Balancer | Application Gateway; Front Door, globally |
| Google Cloud | Network load balancing | Application load balancing |
Global versus regional. Some layer 7 services accept traffic at edge locations worldwide and send it to the nearest healthy region. Examples are Azure Front Door, Google's global application load balancer, and AWS CloudFront in front of a regional balancer. That gives one address for users everywhere and failover between regions (question 589).
With Kubernetes, a LoadBalancer Service usually creates a layer 4 load balancer. A gateway or ingress controller behind it then does the layer 7 work. Some cloud controllers instead create a managed layer 7 load balancer directly (question 371).
540. How does cloud DNS work, including private zones and routing policies?
The short answer
Cloud DNS services, such as AWS Route 53, Azure DNS and Google Cloud DNS, host your domain's records on the provider's global network of name servers. Public zones answer for everyone on the internet. Private zones answer only inside chosen networks, for internal names and private endpoints. Some services also offer routing policies, so one name can return different answers depending on the user's location, the health of each endpoint, or set weights, which is used for failover and gradual migrations.
Record types you use every day:
| Record | Maps | Example |
|---|---|---|
| A, AAAA | A name to an IPv4 or IPv6 address | shop.example.com → 20.61.8.4 |
| CNAME | A name to another name | www.shop.example.com → shop.example.com |
| Alias, or apex alias | The bare domain to a cloud resource | example.com → a load balancer |
| TXT | Text, often for verification | Domain ownership for certificates |
| MX | Mail servers | Email delivery |
TTL, time to live, says how long others may cache an answer, in seconds. Lower it, for example to 60, a day before a planned change, so the change spreads quickly.
Routing policies, mainly in Route 53, with similar features in Azure Traffic Manager:
| Policy | Answer depends on | Use for |
|---|---|---|
| Simple | Nothing | Ordinary records |
| Weighted | Set percentages | Moving 10% of users to a new region |
| Latency or geolocation | Where the user is | Sending users to the nearest region |
| Failover | Health checks | Switching to the standby region automatically |
Private zones answer only for networks linked to them, which is how internal names such as orders-db.shop.internal, and private endpoints, resolve inside your network (question 538).
A DNS caveat for failover: clients and resolvers cache answers, sometimes longer than the TTL. DNS failover takes at least the TTL, often more, so it suits disaster recovery better than instant switching (question 589).
541. What is a CDN, and when should you use one?
The short answer
A content delivery network is a worldwide set of edge servers that keep copies of your content close to users. When a user asks for an image, a script or a page, the nearest edge server answers from its copy, instead of every request travelling to your origin servers. Pages load faster, your servers get far less traffic, and many CDNs also absorb attacks and end TLS at the edge. Use one for static files and media, and increasingly in front of whole websites and APIs.
How it works:
- A user in Mumbai requests
https://shop.example.com/assets/app.3f9c.js. - DNS sends them to the CDN's nearest edge server, in Mumbai.
- Cache hit: the edge already has the file, and returns it in a few milliseconds.
- Cache miss: the edge fetches it once from the origin, your storage or servers, perhaps in Europe, keeps a copy, and returns it.
- The next thousand users in Mumbai get the edge's copy.
What controls caching: HTTP headers from your origin, such as Cache-Control: public, max-age=31536000, immutable for versioned files, or Cache-Control: no-store for personal pages, plus rules in the CDN's own settings.
A good habit: versioned file names. Build tools produce names with a hash, such as app.3f9c.js. The file can be cached for a year, and a new release simply uses a new name, so you never wait for old copies to expire.
What else CDNs do:
- End TLS at the edge, near the user, saving round trips.
- Absorb traffic floods, as part of DDoS protection (question 542).
- Run a web application firewall at the edge.
- Run small functions at the edge, for redirects or header changes.
Products: Amazon CloudFront, Azure Front Door, Google Cloud CDN, and independent ones such as Cloudflare, Akamai and Fastly.
542. What are a web application firewall and DDoS protection, and how do they differ?
The short answer
A web application firewall, WAF, inspects HTTP requests and blocks those that look like attacks on the application itself, such as SQL injection, cross-site scripting, or requests from known bad sources, and it can rate-limit abusive clients. DDoS protection defends against floods of traffic meant to overwhelm your network or servers, by absorbing and filtering huge volumes before they reach you. A WAF protects the app's logic, DDoS protection protects its availability, and they usually work together at the edge.
| WAF | DDoS protection | |
|---|---|---|
| Protects against | Malicious requests | Overwhelming volume |
| Looks at | Request content: paths, parameters, headers, bodies | Traffic volume and patterns |
| Typical attacks | SQL injection, cross-site scripting, bots | Floods of packets or requests |
| Works at | Layer 7 | Layers 3 and 4, and layer 7 floods |
| Examples | AWS WAF, Azure WAF on Front Door or Application Gateway | AWS Shield, Azure DDoS Protection |
A DDoS attack, distributed denial of service, uses thousands or millions of machines, often hijacked home devices, to send so much traffic that real users cannot get through. Cloud providers absorb floods across their huge global networks. Basic protection is included for free. Paid tiers add tuned protection, attack reports and support during an attack.
WAF rules usually start from managed rule sets, such as the OWASP Core Rule Set, which covers the most common web attack patterns. Then you add your own rules: rate limits per client, blocking certain countries if you have no customers there, or allowing only your own paths.
Start in detection mode. A WAF can block legitimate traffic, for example a product description that happens to look like SQL. Run new rules in "log only" mode first, check what they would block, then switch to blocking.
Neither replaces secure code. A WAF catches known patterns. It does not fix an injection flaw in your code, it only makes it harder to reach (question 214).
543. How do you securely reach private machines without exposing SSH or RDP to the internet?
The short answer
Use the cloud's managed access service instead of opening ports: AWS Systems Manager Session Manager, Azure Bastion, or Google's Identity-Aware Proxy. You log in with your normal cloud identity, and they give you a shell or remote desktop through the cloud's control plane. No machine needs a public address or an open SSH or RDP port, and every session is logged. If a bastion host is used instead, keep it locked down, behind single sign-on and a VPN.
Why open SSH and RDP ports are dangerous. Ports 22 and 3389 open to the internet are scanned constantly, within minutes of a machine appearing. Weak passwords, leaked keys and unpatched remote desktop flaws are some of the most common ways in.
| Option | How it works | Open inbound port needed |
|---|---|---|
| AWS Session Manager | An agent on the machine connects out to AWS; you connect through the AWS console or CLI | None |
| Azure Bastion | A managed service in your VNet; you connect through the Azure portal or CLI | None on the machines |
| Google Identity-Aware Proxy | Google checks your identity, then forwards the connection | None from the internet |
| A bastion, or jump host | One hardened machine you SSH into first, then onwards | One, ideally limited to your VPN |
What the managed options give you:
- No public addresses on the machines.
- Your normal identity, with multi-factor authentication, instead of shared keys.
- Permissions controlling who may connect to which machines.
- Audit logs, and often full session recordings.
And in a world of containers: ideally, nobody logs in to machines at all. Configuration comes from code, problems are investigated through logs, metrics and debug containers (question 414), and broken machines are replaced, not repaired.
544. A virtual machine in a private subnet cannot reach the internet; how would you troubleshoot it?
The short answer
Follow the path out, one hop at a time. Check the subnet's route table has a route for internet traffic to a NAT gateway, not to nothing or to an internet gateway the machine cannot use. Check the NAT gateway exists, is healthy, and in AWS sits in a public subnet. Check outbound firewall rules on the machine's security group and subnet rules, and any central firewall in a hub. Check DNS resolves. And check the machine itself: its own firewall and proxy settings.
Step by step:
- Is it DNS or the network? From the machine:
bash
nslookup api.stripe.com # name resolution
curl -sv --max-time 5 https://1.1.1.1 -o /dev/null # a known address, no DNS- The route table. The private subnet needs
0.0.0.0/0pointing at a NAT gateway, or, in a hub-and-spoke design, at the central firewall. A route to an internet gateway does not help a machine with no public address. - The NAT gateway. In AWS, it must be in a public subnet, whose own route table points at the internet gateway. It must have a public address and be in the available state. In Azure, the NAT gateway must be associated with the machine's subnet.
- Outbound firewall rules. Security groups usually allow all outbound traffic by default, but locked-down environments often restrict it. Check network ACLs too, including return traffic, since they are stateless (question 534).
- A central firewall. In hub-and-spoke networks, traffic goes through a firewall that may only allow listed destinations. Its logs show the denied connection.
- The machine itself. Its own firewall, and proxy settings: in many companies, outbound traffic must go through an HTTP proxy.
- The destination. It may block your NAT gateway's address, or require it to be on an allow-list.
Tools that save time: AWS Reachability Analyzer and Azure Network Watcher's connection troubleshoot and next-hop tools trace the path through route tables and firewall rules for you, and say exactly which rule or route blocks it.
545. How would you design a secure network for a three-tier web application in the cloud?
The short answer
Split the network into tiers, each in its own subnets spread across availability zones. The edge tier holds only the public entry point: a CDN with a WAF, and a load balancer. The application tier runs in private subnets, reachable only from the load balancer, with outbound internet through NAT. The data tier runs in its own private subnets, reachable only from the application tier, and managed services are reached through private endpoints. Firewall rules allow only the traffic each tier needs, and administration goes through a managed access service.
The design, tier by tier:
- Edge: a CDN with a WAF and DDoS protection in front, so attacks are filtered before reaching your network (questions 541 and 542).
- Public subnets, one per zone: only the load balancer and the NAT gateways. The load balancer accepts HTTPS, ideally only from the CDN's addresses.
- Private application subnets, one per zone: application servers or Kubernetes nodes, with no public addresses, accepting traffic only from the load balancer's security group. Outbound internet access through NAT, or better, through a firewall with an allow-list.
- Private data subnets: a managed database with a standby in another zone, accepting connections only from the application tier's group, with no internet route at all.
- Private endpoints for storage, the secret store and the container registry, with public access disabled on each (question 538).
Around it:
- Administration through Session Manager or Azure Bastion, never open SSH or RDP (question 543).
- Flow logs on the network, recording accepted and rejected connections, for investigations.
- Infrastructure as code for the whole network, so it is reviewed and repeatable (question 595).
- Two zones at least, three for higher availability, in every tier (question 520).