Senior Load Balancing & Networking Interview Questions (50 MCQs)

Question 1

Fifty independent LB nodes, each with a local, slightly stale view of load, all run least connections. Why does power of two choices often perform better?

Accepted Answer

Least connections makes all nodes pick the same seemingly idle server at once (herding); p2c's randomness breaks this

Answer

P2c synchronizes a complete, real-time view of all servers' load across nodes

Answer

Least connections is technically impossible to implement in a distributed setting

Answer

P2c restricts all requests to two fixed servers, raising cache hits

Question 2

You route by `tenant_id` with consistent hashing for cache locality, but one huge tenant melts a node. Most suitable fix?

Accepted Answer

Consistent hashing with bounded loads: cap each node's load and spill to the next ring node

Answer

Increase virtual nodes so that tenant's load is split automatically

Answer

Switch to `hash(tenant_id) % N` for a more even distribution

Answer

Lower the DNS TTL so clients spread out faster

Question 3

Why does Maglev hashing suit an L4 load balancer tier with many independent nodes?

Accepted Answer

Near-even distribution via a lookup table and minimal disruption, with no shared state between nodes

Answer

It guarantees zero flow movement even when backends are removed

Answer

It enables header- and cookie-based routing at Layer 4

Answer

It eliminates the need for backend health checks

Question 4

How does rendezvous hashing (HRW) work, and what is its trade-off versus ring-based consistent hashing?

Accepted Answer

Compute `hash(key, node)` per node and pick the max; no ring or vnodes, but lookup is `O(N)`

Answer

The key is hashed onto a ring and the next clockwise node chosen; lookup is `O(1)`

Answer

Nodes take keys in turns; no hash computation is needed

Answer

A central coordinator stores key-to-node mappings; it requires network calls

Question 5

You scale a gRPC service from 3 to 10 pods with client-side LB, yet the 7 new pods get almost no traffic. What is an effective server-side fix?

Accepted Answer

Set `MaxConnectionAge` so the server sends `GOAWAY`, forcing clients to reconnect and re-resolve

Answer

Raise `MaxConcurrentStreams` so old pods accept more requests

Answer

Disable keepalive so connections are never closed

Answer

Increase the readiness probe delay on new pods

Question 6

In Envoy, what happens when a cluster's healthy host percentage drops below the panic threshold (default 50%), and why?

Accepted Answer

Traffic goes to all hosts, including unhealthy ones, so the few remaining hosts are not crushed

Answer

All requests are immediately rejected with `503` so the system can recover

Answer

Envoy automatically restarts the unhealthy hosts

Answer

Traffic is sent to only one random healthy host

Question 7

Why is outlier detection usually capped with `max_ejection_percent` (e.g. 10%)?

Accepted Answer

So a systemic issue (like a shared dependency failure) does not eject the whole fleet and cause an outage

Answer

To keep Envoy's memory use for the host list low

Answer

Because Envoy technically cannot track more than 10% of hosts

Answer

So ejected hosts get restarted faster

Question 8

During a traffic spike, overloaded servers time out on health checks, the LB removes them, and the rest fall one after another. The most fundamental fix?

Accepted Answer

A cheap, prioritized health check, and shedding load with `503` instead of failing health under load

Answer

Raise the health check interval to several minutes

Answer

Remove health checks entirely and rely on round robin

Answer

Switch from least connections to IP hash

Question 9

A request path has 4 layers and each makes up to 3 attempts. When the bottom service slows down, what happens and what is the fix?

Accepted Answer

Load amplifies up to `3^4 = 81`x; retry at one layer only, with a retry budget and backoff

Answer

Load at most triples; no problem since retries are always safe

Answer

Load grows 12x; remove timeouts in every layer

Answer

No amplification occurs because the LB deduplicates retries

Question 10

A reverse proxy talking to a backend on one IP:port starts failing with `EADDRNOTAVAIL` under high load. Cause and correct fixes?

Accepted Answer

Ephemeral port exhaustion for that destination tuple; connection pooling, more source IPs, or more destination IP:ports

Answer

The proxy's disk is full; delete logs

Answer

The backend certificate expired; renew it

Answer

DNS cannot resolve; lower the TTL

Question 11

Despite graceful shutdown in the app, every Kubernetes rolling update causes a few seconds of `502`s. Most likely cause and fix?

Accepted Answer

Endpoint removal reaches kube-proxy/ingress late; add a few-second `preStop` sleep and keep serving after `SIGTERM`

Answer

The readiness probe is too fast; disable it entirely

Answer

Replica count is too low; set `maxSurge` to zero

Answer

Cluster DNS is slow; increase the TTL

Question 12

Why is anycast risky for long-lived TCP connections (like multi-hour downloads)?

Accepted Answer

A mid-connection BGP route change sends packets to another PoP with no TCP state, which resets it

Answer

Anycast only supports UDP, so TCP does not work at all

Answer

Anycast automatically closes connections after 60 seconds

Answer

Anycast picks a random destination IP for every packet

Question 13

In a keepalived pair, the heartbeat link fails while both nodes are healthy. What is the risk and how is it mitigated?

Accepted Answer

Split-brain: both claim the VIP; use redundant heartbeat paths, fencing, or a witness/quorum

Answer

Both nodes shut down; increase the interval

Answer

The VIP is deleted forever; repoint DNS to another IP

Answer

No risk, since VRRP always has exactly one master

Question 14

Your L4 LB tier sits behind a router doing ECMP. Adding an LB node resets some live connections. What design minimizes this?

Accepted Answer

All nodes pick backends via consistent hashing (like Maglev), plus local connection tracking

Answer

Each LB node uses independent round robin with a local counter

Answer

Replace ECMP with DNS round robin across LB nodes

Answer

Raise all connections' idle timeout to one day

Question 15

In Layer 2 DSR, why must backends hold the VIP on loopback but not answer ARP for it?

Accepted Answer

So they accept VIP-addressed packets and reply from the VIP without stealing traffic from the LB

Answer

So traffic between backends is encrypted

Answer

So the LB can modify HTTP headers

Answer

So health checks can run over ARP

Question 16

TLS session resumption on your 20-node L7 fleet is near zero and handshake CPU is high. Likely cause?

Accepted Answer

Each node has its own session ticket key, so clients landing on another node cannot resume

Answer

TLS 1.3 does not support session resumption

Answer

The certificate is a wildcard, which disables resumption

Answer

HTTP/2 always does a full handshake per request

Question 17

You scaled WebSocket servers from 4 to 8, but the 4 old ones stay hot and the new ones idle. What actually rebalances load?

Accepted Answer

Hot servers gradually close a fraction of connections with jitter so clients reconnect onto new servers

Answer

Switch the LB algorithm to round robin for existing connections

Answer

Lower the DNS record's TTL to 5 seconds

Answer

Enable cross-zone load balancing on the LB

Question 18

A legacy app relies on sticky sessions and has autoscaling. Which problems should you expect?

Accepted Answer

New instances only get new sessions, and scale-in destroys users' sessions

Answer

Autoscaling is fully disabled because the LB forbids it

Answer

All users go to a random new server on each request

Answer

Sticky sessions make all instances share one session

Question 19

How do hedged requests reduce tail latency, and what constrains them?

Accepted Answer

If no reply by about p95, a second request goes to another replica; only for idempotent ops with capped extra load

Answer

Every request is sent to three replicas upfront; safe for all operations

Answer

Slow requests are queued until the server is idle

Answer

Every layer's timeout is cut to p50

Question 20

During a flash sale, capacity is 3x short. What is the best behavior at the edge/LB layer?

Accepted Answer

Priority load shedding: quickly reject low-value traffic with `503` and `Retry-After`, protecting checkout

Answer

Accept every request and queue it until it can be processed

Answer

Randomly reject 66% of requests regardless of type

Answer

Turn off health checks so servers never leave rotation

Question 21

Zone-aware routing (like Envoy) keeps traffic in the local AZ. When must it spill over to other zones?

Accepted Answer

When the local zone's healthy hosts are insufficient for that zone's share of traffic

Answer

Whenever request latency exceeds 1 ms

Answer

Never; traffic must always stay in the same zone

Answer

Only when the entire region has failed

Question 22

Your fleet mixes several instance generations and has noisy neighbors; static weighted round robin performs poorly. Better approach?

Accepted Answer

Dynamic weights based on backend-reported load (e.g. CPU or QPS via ORCA)

Answer

Hand-calculating weights once a year from hardware specs

Answer

Using IP hash so each client stays on one instance

Answer

Removing every instance older than one generation

Question 23

Your API gateway validates JWTs, so a team proposes services skip all security checks. What is the core problem?

Accepted Answer

Internal traffic can bypass the gateway, and object-level authorization can only happen in the service

Answer

The gateway technically cannot validate JWTs

Answer

Validating JWTs at the gateway increases DNS TTL

Answer

Services without security checks always get slower

Question 24

A global limit of 1000 req/s per API key across 20 autoscaling gateway nodes. Which approach has a sensible trade-off?

Accepted Answer

Local token buckets with periodic sync to a central store; approximate accuracy for low latency

Answer

A fixed 50 per node; always accurate even as node count changes

Answer

A synchronous Redis call per request; no latency cost or dependency

Answer

Enforce the limit only on the client, since clients are always trustworthy

Question 25

The current `A` record has a TTL of 86400 and you migrate to a new LB tomorrow at 10:00. What is the right plan?

Accepted Answer

Lower the TTL to e.g. 60 at least 24 h ahead, then switch, and keep the old LB running a while

Answer

At 10:00 set TTL to 60 and change the IP together; everyone moves within a minute

Answer

Increase the TTL so caches are more stable, then change the IP

Answer

Shut down the old LB exactly when changing the record

Question 26

Asian users on a public resolver are being sent to the European datacenter by GeoDNS. Likely cause and remedy?

Accepted Answer

GeoDNS sees the resolver's location, not the user's; support EDNS Client Subnet or use anycast

Answer

The GeoIP database lists Asia as Europe; raise the TTL

Answer

Asian users use IPv6, which GeoDNS does not support

Answer

The Asian datacenter has no `MX` record

Question 27

The LB's control plane (config API and discovery) is down for an hour. What does a statically stable design expect from the data plane?

Accepted Answer

Keep serving with the last known-good config and endpoints, independent of the control plane

Answer

Halt all traffic until the control plane returns and confirms config

Answer

Wipe config and revert to factory defaults

Answer

Route all traffic to one fixed backend until fresh data arrives

Question 28

Region A was down for 2 hours and traffic moved to B. A now reports healthy. Why shouldn't GSLB instantly move 100% of A's traffic back?

Accepted Answer

A's caches are cold, its autoscaling shrank, and pools are empty; a sudden shift can knock it over again

Answer

Because DNS does not allow record changes more than once a day

Answer

Because B's traffic is automatically locked there forever

Answer

Because region A's health checks always give false results

Question 29

The received header is `X-Forwarded-For: 1.2.3.4, 203.0.113.7, 10.0.0.5`, and only your LB (`10.0.0.5`) is trusted. Which is the real client IP?

Accepted Answer

`203.0.113.7`, the first IP from the right not belonging to trusted proxies

Answer

`1.2.3.4`, since the leftmost IP is always the original client

Answer

`10.0.0.5`, since the last IP is always the client

Answer

None; `X-Forwarded-For` must never be read

Question 30

What causes HTTP request smuggling between a reverse proxy and a backend?

Accepted Answer

Differing interpretation of request boundaries (e.g. `Content-Length` vs `Transfer-Encoding`) on a shared connection

Answer

Using self-signed certificates between proxy and backend

Answer

Having gzip enabled on backend responses

Answer

Setting a low TTL on the service's DNS record

Question 31

Why is 'least outstanding requests' more meaningful than 'least connections' for an L7 proxy talking HTTP/2 to backends?

Accepted Answer

With multiplexing, each backend may have one connection, so connection count does not reflect load

Answer

HTTP/2 does not let the proxy count connections

Answer

Least connections is only defined for UDP traffic

Answer

Least outstanding requests needs no backend info, but least connections does

Question 32

During a regional failover, GSLB moves users to region B, whose DB replica lags by 30 s. Which consequence must be explicitly designed for?

Accepted Answer

Recent writes lost or not visible (RPO) and broken read-your-writes

Answer

Larger TLS certificate sizes in region B

Answer

Automatic disabling of rate limiting in region B

Answer

Conversion of all connections to UDP

Question 33

What benefit does shuffle sharding provide at the LB or worker pool layer?

Accepted Answer

Each customer maps to a small random combination of nodes, so a toxic customer affects only a tiny fraction of others

Answer

All customers spread across all nodes so load is perfectly even

Answer

Each customer gets a dedicated node and no node is shared

Answer

Shards are randomly reshuffled every hour for security

Question 34

In Kubernetes, what trade-off does `externalTrafficPolicy: Local` on a LoadBalancer Service bring?

Accepted Answer

Client IP is preserved and the extra hop removed, but load spreads per node and pods may become imbalanced

Answer

All pods are placed on one node, reducing availability

Answer

TLS is automatically terminated on every node

Answer

Traffic becomes reachable only from inside the cluster

Question 35

A massive SYN flood hits your VIP. Why does terminating TCP at an edge LB/proxy with SYN cookies help?

Accepted Answer

No state is held until the handshake completes, and backends only see completed connections

Answer

SYN cookies automatically block the attacker's IP in BGP

Answer

SYN cookies convert traffic to UDP, which cannot be flooded

Answer

SYN cookies remove the need for TLS, freeing CPU

Question 36

Why is an L4 LB that hashes on the 5-tuple problematic for HTTP/3 (QUIC)?

Accepted Answer

With connection migration the client IP/port changes; the LB must route by Connection ID

Answer

QUIC runs over TCP, so the LB cannot see the port

Answer

HTTP/3 always uses a random port, making hashing meaningless

Answer

QUIC broadcasts packets to all servers

Question 37

What trade-off does proxyless gRPC with xDS have versus a sidecar-based mesh?

Accepted Answer

Removes the sidecar hop and cost, but only for supported gRPC clients and tied to library versions

Answer

Higher latency, but supports every protocol like Redis and MySQL

Answer

No control plane needed, but no load balancing

Answer

More security, but full incompatibility with Kubernetes

Question 38

Why is an AP-style registry (like Eureka) often preferred over CP for service discovery?

Accepted Answer

During a partition, a slightly stale endpoint list beats no answer, and clients tolerate errors with retries

Answer

AP registries always have more accurate, fresher data than CP

Answer

CP registries technically cannot receive heartbeats

Answer

AP registries need no replication

Question 39

Behind an ALB with a 60 s idle timeout, a Node.js backend with the default 5 s `keepAliveTimeout` emits sporadic `502`s. Cause and fix?

Accepted Answer

The backend closes idle connections first while the LB sends on them; make backend keep-alive exceed the LB idle timeout

Answer

Node.js does not support HTTP/1.1; enable HTTP/2

Answer

The LB idle timeout is too short; reduce it to 1 second

Answer

The certificate between ALB and backend has expired

Question 40

Why is mesh mTLS stronger than IP-based NetworkPolicy for service-to-service authorization?

Accepted Answer

A cryptographic workload identity (like a SPIFFE ID) is independent of churning pod IPs, so policy targets services

Answer

mTLS makes traffic faster than unencrypted traffic

Answer

mTLS eliminates the need for certificates entirely

Answer

NetworkPolicy is technically unsupported in Kubernetes

Question 41

Why do large providers typically layer anycast/ECMP → L4 LB → L7 proxy → services?

Accepted Answer

L4 is cheap and near-stateless, letting the expensive L7 tier (TLS, routing) scale horizontally and independently

Answer

Because each layer is a full backup of the previous one and repeats its work

Answer

Because L7 proxies can never have more than one instance

Answer

Because anycast only works with L7 proxies and bypasses L4

Question 42

With weights A=5, B=1, C=1, why is smooth weighted round robin (as in NGINX) better than the naive version?

Accepted Answer

It yields an order like `AABACAA` instead of `AAAAABC`, eliminating bursts on A

Answer

It automatically adjusts weights based on servers' CPU

Answer

It raises C's share to match A for fairness

Answer

It only works when all weights are equal

Question 43

When draining a proxy, how do you move HTTP/2 and HTTP/1.1 keep-alive clients to another instance without abrupt cuts?

Accepted Answer

Send a `GOAWAY` frame for HTTP/2 and `Connection: close` on the next HTTP/1.1 response

Answer

Send RST on all connections so clients reconnect immediately

Answer

Change the DNS record and wait for TTL expiry without touching connections

Answer

Drop `MaxConcurrentStreams` to zero without notifying clients

Question 44

You expect a live event to spike traffic 20x within 30 seconds. What is an overlooked risk in a managed cloud load balancer?

Accepted Answer

The LB itself scales gradually and may error before scaling; pre-warming or load testing is needed

Answer

Managed LBs have unlimited, instant capacity, so there is no risk

Answer

The LB automatically disables HTTPS under high traffic

Answer

The LB sends all traffic to one AZ during spikes

Question 45

In a mesh with 2000 sidecars and a service with 500 instances, what problem do active health checks from every sidecar cause?

Accepted Answer

Millions of probes per interval; better to rely on centralized health from the control plane plus outlier detection

Answer

No problem, since probes are extremely lightweight

Answer

Sidecars cannot probe more than 10 instances

Answer

Probes cause mTLS certificates to expire

Question 46

5000 clients each hold persistent connections to all 1000 backends, and backend memory is consumed by connections. Suitable technique?

Accepted Answer

Deterministic subsetting: each client connects only to a balanced subset of backends

Answer

Raise the idle timeout so connections churn less

Answer

Switch all clients to IP hash onto one backend

Answer

Drop keep-alive and open a new connection per request

Question 47

What property does peak EWMA (in Finagle and Linkerd) have over a simple latency average?

Accepted Answer

It reacts instantly to latency spikes but decays cautiously, and is combined with p2c

Answer

It keeps only each backend's all-time peak latency forever

Answer

It ignores latency and only counts connections

Answer

It sorts all backends by latency every second

Question 48

A service runs active-active across 3 regions with equal traffic. To survive losing one region, roughly what is the max normal utilization per region?

Accepted Answer

About 66% (minus headroom), since each surviving region must take 1.5x load

Answer

About 90%, since GSLB manages the extra load

Answer

About 33%, since each region must take all traffic

Answer

About 50%, since half of capacity must always be free

Question 49

Your LB routes between tenants by SNI in TLS passthrough mode. What must change once clients enable Encrypted Client Hello (ECH)?

Accepted Answer

The real SNI is encrypted and only the outer public name is visible; the LB must hold the ECH key (client-facing server)

Answer

Nothing changes, since ECH only encrypts the HTTP request body

Answer

The LB must route by each tenant's destination IP instead, needing no keys

Answer

ECH downgrades the TLS connection to plain unencrypted TCP

Question 50

How does kube-proxy in iptables mode pick a backend, and what does IPVS mode add?

Accepted Answer

iptables picks probabilistically at random per connection; IPVS adds algorithms like least connection and faster lookups

Answer

iptables picks per HTTP request with least request; IPVS has only round robin

Answer

iptables routes by the `Host` header; IPVS by cookie

Answer

iptables always picks the first pod; IPVS the last pod