Question 1

Why do reusable patterns and components matter when building distributed systems?

Accepted Answer

Historically, almost every distributed system was a bespoke, one-off solution &mdash; even when systems shared nearly identical concepts and logic, it was hard to reuse code or apply known designs. This is similar to software development before object-oriented programming made reusable objects and design patterns mainstream.The rise of containers and container orchestrators changed this. Containers are a packaging and deployment boundary that act as the reusable building block for distributed sy

Question 2

What are the trade-offs between a microservices architecture and a monolith?

Accepted Answer

A monolith places all functionality inside a single, tightly-coordinated application. A microservices architecture breaks the system into many small services running in separate processes that communicate over defined APIs.Benefits of microservices:Agility &mdash; each service is small enough for a single "two-pizza" team, and formal APIs decouple teams so they can release on independent schedules.Reliability &mdash; a stable API contract reduces tight synchronization between teams.Independent s

Question 3

What is the sidecar pattern, and when would you use it?

Accepted Answer

The sidecar pattern is a single-node pattern made of two co-scheduled containers in one atomic group (e.g., a Kubernetes Pod): the application container holds the core logic, and the sidecar container augments or improves it &mdash; often without the application container's knowledge. They share resources such as parts of the filesystem, network, and other namespaces.Common uses:Adding capabilities to legacy apps &mdash; e.g., an old service only speaks HTTP. Bind it to localhost and add an ngin

Question 4

What is the ambassador pattern, and how does it differ from a sidecar?

Accepted Answer

In the ambassador pattern, an ambassador container brokers interactions between the application container and the outside world. Like the sidecar, the two containers are tightly paired on a single machine, but the focus is on mediating outbound communication rather than augmenting the app itself.Canonical example &mdash; client-side sharding: the application simply connects to what looks like a single storage backend on localhost. In reality that is a sharding-proxy ambassador (e.g., twemproxy f

Question 5

How can an ambassador be used for experimentation or request splitting?

Accepted Answer

Request splitting sends some fraction of traffic to an alternative implementation of a service. The application connects to localhost as usual, and the ambassador decides where each request actually goes.Two main uses:A/B experiments &mdash; route, say, 90% of traffic to the production service and 10% to a new beta version to compare reliability and performance (e.g., nginx upstream with weights).Tee / shadow traffic &mdash; send a copy of all traffic to a new, undeployed version; return only th

Question 6

What is the adapter pattern, and what problem does it solve?

Accepted Answer

Real applications are heterogeneous: some code is written in-house, some comes from vendors, and some is off-the-shelf binaries &mdash; in many languages, each with its own conventions for logging, metrics, and health checks. To operate them, you need common interfaces.In the adapter pattern, an adapter container modifies the application container's interface to conform to a predefined standard expected across all applications. The application keeps its native behavior; the adapter translates it

Question 7

Why do you need at least two replicas for a highly available stateless service?

Accepted Answer

A replicated load-balanced service runs identical stateless servers behind a load balancer (round-robin or session-sticky). Even for the smallest service you need at least two replicas to claim a highly-available SLA.Why the math forces this: a three-nines (99.9%) SLA permits only about 1.4 minutes of downtime per day (24 &times; 60 &times; 0.001). With a single instance, every software rollout or crash is downtime &mdash; you'd have to complete an upgrade in under 1.4 minutes per day, and if yo

Question 8

What is the difference between a readiness probe and a liveness (health) probe?

Accepted Answer

Both probes are health signals used in a replicated service, but they answer different questions:Liveness / health probe &mdash; answers "is this container broken and in need of a restart?" The orchestrator uses it to restart a hung or dead container.Readiness probe &mdash; answers "is this instance ready to serve user requests right now?" The load balancer uses it to decide whether to send traffic to the instance.The distinction matters because many applications are alive but not yet ready: the

Question 9

How is session affinity maintained in a load-balanced service, and why use consistent hashing for it?

Accepted Answer

Some services need a user's requests to keep landing on the same replica (session tracking). This is commonly done by hashing the source and destination IP addresses to pick a replica &mdash; as long as those stay constant, requests stick to one server.Caveats: IP-based affinity works inside a cluster but breaks across the internet because of NAT, so for external clients you should track sessions with a cookie or HTTP header instead. Also, inserting a cache tier can break IP affinity, because th

Question 10

Why deploy a cache as a separate replicated tier rather than as a sidecar next to each web server?

Accepted Answer

A caching web proxy (e.g., Varnish) sits between users and a stateless web tier; if two users request the same page, only one request reaches the backend and the second is served from memory. The question is how to deploy it.Sidecar cache (simplest): co-locate a cache with each web server. The drawback is that you are forced to scale the cache 1:1 with the web tier, and every replica caches the same pages, so the same page is stored many times &mdash; wasting memory and lowering the hit rate.Sep

Patterns and Containers(2 questions)

Why do reusable patterns and components matter when building distributed systems?

What are the trade-offs between a microservices architecture and a monolith?

Single-Node Patterns(4 questions)

What is the sidecar pattern, and when would you use it?

What is the ambassador pattern, and how does it differ from a sidecar?

How can an ambassador be used for experimentation or request splitting?

What is the adapter pattern, and what problem does it solve?

Replicated and Cached Serving(6 questions)

Why do you need at least two replicas for a highly available stateless service?

What is the difference between a readiness probe and a liveness (health) probe?

How is session affinity maintained in a load-balanced service, and why use consistent hashing for it?

Why deploy a cache as a separate replicated tier rather than as a sidecar next to each web server?

How does a cache's hit rate affect overall system capacity, and what is the risk of a cache failure?

Beyond caching, what does an edge / reverse-proxy layer typically provide?

Sharded Serving(5 questions)

How do sharded services differ from replicated services, and when should you shard?

Why does a sharded cache use memory far more efficiently than a replicated cache?

What properties must a good sharding function have, and how do you choose the shard key?

What is consistent hashing, and why is it essential when changing the number of shards?

What is a "hot shard," and how do you handle one?

Scatter/Gather(3 questions)

What is the scatter/gather pattern, and what kind of scaling does it provide?

What is the "straggler problem" in scatter/gather, and why does tail latency dominate?

Why doesn't adding more leaves always make a scatter/gather system faster?

Functions and Event-Driven Processing(4 questions)

When does FaaS / event-driven processing make sense, and when does it not?

What are the main operational challenges of building systems with FaaS?

What is the decorator pattern in FaaS?

What is an event-driven pipeline, and how does it differ from a microservices architecture?

Ownership and Master Election(3 questions)

When can you avoid master election and just run a singleton?

How is distributed master election (or locking) typically implemented?

Why do distributed locks need TTLs and fencing (resource versions), and what bug does fencing prevent?

Batch Computational Patterns(5 questions)

What is a work queue system, and how is it factored into reusable interfaces?

How do you decide when to scale a work queue's worker parallelism?

Describe the event-driven batch patterns used to link work queues: copier, filter, splitter, sharder, and merger.

What role does a publisher/subscriber system like Kafka play, and what do partitions and replication factor control?

Compare the join (barrier synchronization) and reduce coordinated-batch patterns.