Networking & Protocols
Proxy and load balancing: multiple-choice review
Crux Multiple-choice synthesis across the proxy and load-balancing unit — algorithm tradeoffs, L4 vs L7, health detection, sticky-session anti-patterns, retry storms, and resilient topology.
Your altitude — climbing toward senior
ZeroJuniorMiddleSenior
You are at senior altitude — in orbitSix questions that cut across the whole unit. Each one mirrors a decision you make on a real platform — not a definition to recite, but a tradeoff to weigh between speed, fairness, and state under load.
Goal
Confirm you can connect algorithm choice, L4 vs L7, health detection, session strategy, and cascade control — the synthesis the individual lessons built toward.
Quiz
Completed
A fleet of 50 backends serves requests whose cost varies wildly (1 ms API calls mixed with 200 ms uploads). Which balancing algorithm is the right default, and why?
Heads-up Round-robin distributes request COUNT evenly, not load. A 200 ms upload and a 1 ms call each consume one turn, so a backend handling uploads falls behind despite getting its fair share.
Heads-up Least-connections adapts to load but herds: when a recovered backend's count drops, every client sees the same minimum at once and stampedes it. It also needs an O(log N) scan per request.
Heads-up IP-hash is a session-affinity tool, not a load balancer. It clusters badly behind NAT and ignores real-time load entirely.
Quiz
Completed
You must route /api/payments/ to a PCI-compliant pool and /api/catalog/ to a general pool, retry on HTTP 5xx, and emit per-endpoint metrics. Which layer is required and what does it cost?
Heads-up L4 sees only the IP:port 4-tuple. It cannot read the URL path or distinguish /payments/ from /catalog/, and it does not understand HTTP 5xx to retry on it.
Heads-up L4 never parses HTTP headers; only L7 does. Content routing and 5xx retries are L7-only capabilities.
Heads-up DNS resolves hostnames, not URL paths. Path routing requires an L7 proxy that parses the request line.
Quiz
Completed
A backend's disk fills; it still accepts TCP connections but returns 500 on every HTTP request. Your active health check is a TCP SYN probe and reports the backend healthy. What is the correct fix?
Heads-up A faster TCP SYN probe still only checks that TCP connects. The process IS accepting connections, so it reports healthy no matter how often you probe.
Heads-up Least-loaded routing reduces traffic to a slow node but never removes a node returning instant 500s — its connections close fast, so it looks idle and attracts MORE traffic.
Heads-up A stub that always returns 200 is the dangerous opposite of the fix — it masks the broken application and keeps routing traffic into the 500 storm.
Quiz
Completed
A team enables sticky sessions to fix users getting logged out when requests land on different backends. After deploy, load imbalance worsens and a crashed node logs out thousands of users. How do you read this?
Heads-up A longer cookie TTL pins clients even harder, worsening imbalance and making session loss on crash more likely, not less. The pinning itself is the problem.
Heads-up IP-hash is still affinity — it keeps the SPOF and adds NAT clustering and breakage on IP change. Swapping affinity flavors does not remove the stateful-backend design flaw.
Heads-up The imbalance and the mass logout are direct consequences of pinning. Sticky sessions reduce resilience by creating a session single point of failure.
Quiz
Completed
A backend has a 500 ms GC pause. Clients and three intermediate service layers each retry once on timeout. A 500 ms blip becomes a cluster-wide outage. What is the minimum set of mitigations?
Heads-up More retries amplify overload further — across K layers retries compound toward 2^K× load. Retrying harder into a drowning backend is exactly what causes the storm.
Heads-up Capacity delays the collapse but retry amplification scales with load too; the new nodes get pulled into the same storm. You must bound retries, not just add capacity.
Heads-up Removing jitter synchronizes every timed-out client to retry at the same instant — a thundering herd. Jitter exists precisely to spread the burst out.
Quiz
Completed
Your single L7 load balancer is itself the single point of failure: when it crashes, all traffic drops. What topology removes the SPOF, and what constraint does it impose?
Heads-up Active-passive failover takes 10–30 s to detect and promote, long enough for users to notice. Anycast + ECMP converges in under a second across an active cluster.
Heads-up DNS TTLs of 60–300 s mean clients keep hitting the dead LB until the cache expires; DNS has no per-request health awareness. DNS does geo-routing, not LB failover.
Heads-up Vertical scaling does not eliminate a single point of failure — the machine, NIC, rack, or power can still fail. Redundancy across machines is the only fix.
Recap
The through-line of the unit is one tradeoff triangle — speed, fairness, and state — resolved at every layer. Power-of-two-choices is the fair O(1) default; L7 buys content routing and 5xx retries at 100× the L4 latency; active plus passive health checks cover both crashes and silent 5xx; sticky sessions are a stateful-backend workaround that Redis replaces; retry budgets, circuit breakers, and jitter bound the cascade; and anycast + BGP ECMP removes the LB’s own single point of failure at the cost of statelessness.