CTO Systems Architecture & Infrastructure 5 — Questions and Answers
Question 1: A CTO must design a system that processes 10TB of clickstream data daily for real-time personalization. Which lambda architecture component handles low-latency serving of precomputed batch results while the speed layer handles recency?
- Speed layer (stream processor)
- Serving layer backed by a fast read store (e.g., Cassandra, Redis) (Correct answer)
- Batch layer (Hadoop/Spark jobs)
- Message broker (Kafka)
Correct answer: Serving layer backed by a fast read store (e.g., Cassandra, Redis)
The serving layer merges batch views with real-time stream results and exposes them via a low-latency read store for online queries.
Question 2: Which infrastructure security control most effectively prevents a compromised container from accessing credentials intended for a different service in the same Kubernetes cluster?
- Running all containers as non-root
- Workload identity with service account token projection scoped per pod (Correct answer)
- Using a single shared Kubernetes secret across all namespaces
- Network policies that block inter-namespace TCP traffic
Correct answer: Workload identity with service account token projection scoped per pod
Service account tokens projected into individual pods allow fine-grained IAM role binding per workload, so a compromised pod can only access its own credentials.
Question 3: A CTO is assessing a proposal to migrate from a 3-tier architecture to an event-sourcing model. What is the most significant operational complexity introduced by event sourcing?
- Higher write latency on every event
- The need for event schema versioning and handling of old event formats during replay (Correct answer)
- Inability to support CQRS read models
- Lack of support from major cloud providers
Correct answer: The need for event schema versioning and handling of old event formats during replay
As systems evolve, old events in the log must still be replayable, requiring disciplined schema evolution strategies (e.g., upcasting, versioned event types).
Question 4: When evaluating a content delivery network for a SaaS application, which metric best indicates CDN effectiveness for dynamic, uncacheable API traffic?
- Cache hit ratio
- Anycast routing convergence time and TCP connection reuse (persistent keep-alive to origin) (Correct answer)
- Number of PoP locations globally
- Supported TLS cipher suites
Correct answer: Anycast routing convergence time and TCP connection reuse (persistent keep-alive to origin)
For dynamic content, CDN value comes from protocol optimization—anycast routing to the nearest PoP and persistent TCP connections to origin—not caching.
Question 5: A CTO is selecting a consensus algorithm for a distributed configuration store that must tolerate node failures without split-brain. Which algorithm is used by etcd (the Kubernetes backing store)?
- Paxos
- Raft (Correct answer)
- Zab (ZooKeeper Atomic Broadcast)
- PBFT (Practical Byzantine Fault Tolerance)
Correct answer: Raft
etcd uses the Raft consensus algorithm, which was designed for understandability and provides strong leader-based consistency with tolerated minority node failures.
Question 6: Which observability signal is most useful for diagnosing why a specific user's request was slow when distributed tracing is not yet implemented?
- Aggregate p99 latency from Prometheus metrics
- Structured log correlation across services using a shared request ID propagated in headers (Correct answer)
- CPU flame graphs sampled at a 1-second interval
- Error rate dashboards in Grafana
Correct answer: Structured log correlation across services using a shared request ID propagated in headers
A shared request ID threaded through structured logs lets engineers reconstruct the per-request execution path across services without a full tracing infrastructure.
Question 7: A CTO is reviewing the architecture of a globally distributed SaaS product. Which consistency model does Google Spanner achieve that makes it unique among distributed databases?
- Eventual consistency with high availability (AP in CAP theorem)
- Externally consistent transactions across geographically distributed nodes using TrueTime (Correct answer)
- Causal consistency using vector clocks per partition
- BASE semantics with tunable read consistency per query
Correct answer: Externally consistent transactions across geographically distributed nodes using TrueTime
Spanner achieves external consistency (a stronger guarantee than serializability) globally by using GPS/atomic clock-based TrueTime to bound clock uncertainty.
CTO Systems Architecture & Infrastructure 5
A CTO must design a system that processes 10TB of clickstream data daily for real-time personalization.
Which lambda architecture component handles low-latency serving of precomputed batch results while the speed layer handles recency?