Top 30 Scaling Interview Questions and Answers

30 questions. Pick an answer, or open “Show the answer” to read it.

Answers are graded in your browser. Nothing is saved, and no XP or streak is earned here. The app keeps score.

1. Question 1 of 30

   What is the primary advantage of employing the Node.js cluster module in a multi-core environment?

   • aIt facilitates the distribution of application load across multiple physical servers using an internal load balancer.
   • bIt provides a low-overhead mechanism for inter-process communication, ideal for tasks requiring frequent data exchange.
   • cIt allows a single Node.js application to utilize all available CPU cores for I/O-bound network operations on one machine.
   • dIt enables direct memory sharing between worker processes for CPU-intensive computations.
   Show the answer

   Answer: c · It allows a single Node.js application to utilize all available CPU cores for I/O-bound network operations on one machine.

   The cluster module's core purpose is to enable a single Node.js application to fully utilize all CPU cores on a *single* multi-core machine for I/O-bound tasks like network applications. It is not designed for direct memory sharing (that's worker_threads), distributing across multiple physical servers, or low-overhead IPC for frequent data exchange, as IPC overhead is noted as high.

   Read the full bite: Node.js Cluster: Scaling on a Single Machine

2. Question 2 of 30

   What is the primary benefit of using Uvicorn workers for a FastAPI application in production?

   • aIt automatically resolves performance issues caused by blocking I/O operations.
   • bIt simplifies debugging by providing clearer error tracebacks.
   • cIt enables automatic code reloading when changes are detected.
   • dIt allows the application to utilize multiple CPU cores for concurrent request processing.
   Show the answer

   Answer: d · It allows the application to utilize multiple CPU cores for concurrent request processing.

   Uvicorn workers are designed to scale your application by running multiple processes, each capable of handling requests, thereby utilizing all available CPU cores to process requests concurrently. Options A and B are incorrect because workers actually complicate debugging and disable auto-reloading; option A is wrong as workers do not fix fundamentally slow or blocking application logic.

   Read the full bite: Uvicorn Workers: Scaling Your FastAPI App

3. Question 3 of 30

   Why does writing millions of tiny objects under one sequential key prefix limit object-store throughput?

   • aEach tiny write incurs request overhead and a single prefix can hotspot one partition
   • bSequential prefixes consume more storage per object
   • cSmall objects bypass the cache and must be written twice
   • dObject stores reject keys that share a common prefix
   Show the answer

   Answer: a · Each tiny write incurs request overhead and a single prefix can hotspot one partition

   Throughput is bounded by request rate plus partition distribution, so per-object overhead and a single hot prefix throttle writes. Prefixes do not increase storage size, are not rejected, and small objects are not written twice.

   Read the full bite: Maximizing object-store throughput for small files

4. Question 4 of 30

   In a cluster setup, what is the primary (master) process responsible for?

   • aForking and supervising worker processes while workers serve traffic
   • bHandling all incoming HTTP requests directly for consistency
   • cRunning the heaviest CPU computation while workers stay idle
   • dSharing a single mutable memory heap across all workers
   Show the answer

   Answer: a · Forking and supervising worker processes while workers serve traffic

   The primary forks workers and restarts them on exit; the workers handle requests across cores. It does not serve traffic itself, share a heap, or run the heavy compute.

   Read the full bite: Scaling across cores with cluster and os

5. Question 5 of 30

   Adding GPU workers yields diminishing throughput gains. What is the most common root cause to investigate first?

   • aThe optimizer learning rate is set incorrectly
   • bThe model has too few parameters to train
   • cGPUs cannot run in parallel by design
   • dGradient synchronization communication overhead growing with worker count and interconnect limits
   Show the answer

   Answer: d · Gradient synchronization communication overhead growing with worker count and interconnect limits

   Data-parallel training all-reduces gradients each step, and that communication cost grows with workers and is bounded by interconnect speed, capping scaling. Learning rate and parameter count do not explain sublinear scaling.

   Read the full bite: Diagnosing poor distributed training scaling

6. Question 6 of 30

   An application's read traffic far exceeds its writes and is slowing the primary. How does a read replica help, and what must the app tolerate?

   • aIt is a synchronous failover standby and serves no live traffic
   • bIt serves read-only queries to offload the primary, but reads may be slightly stale due to lag
   • cIt guarantees reads are always perfectly up to date with the primary
   • dIt accepts writes too, doubling write capacity with no consistency cost
   Show the answer

   Answer: b · It serves read-only queries to offload the primary, but reads may be slightly stale due to lag

   A read replica offloads reads from the primary, scaling read-heavy workloads, but asynchronous replication means replicas can lag and return slightly stale data. It is read-only and not a failover standby.

   Read the full bite: Read replicas in managed relational databases

7. Question 7 of 30

   For which use case is a load balancer most crucial in an AI model serving architecture?

   • aStoring and managing different versions of a machine learning model
   • bPerforming a large, one-time batch inference job on a dataset
   • cDeveloping and testing a model on a local machine with a single user
   • dServing real-time predictions to millions of users with high availability requirements
   Show the answer

   Answer: d · Serving real-time predictions to millions of users with high availability requirements

   The card explicitly states that a load balancer is essential for real-time inference in production environments requiring high availability and fault tolerance. It is unnecessary for offline batch jobs or local development, and model versioning is a separate concern.

   Read the full bite: Load Balancing for Model Serving

8. Question 8 of 30

   What is the primary objective of a Scrum of Scrums (SoS) meeting?

   • aTo conduct in-depth technical design sessions for shared components and services.
   • bTo re-evaluate project timelines and reallocate resources across multiple teams.
   • cTo provide a detailed status update on each team's individual sprint progress.
   • dTo identify and address inter-team dependencies, blockers, and integration risks.
   Show the answer

   Answer: d · To identify and address inter-team dependencies, blockers, and integration risks.

   The core purpose of a Scrum of Scrums is to identify and resolve inter-team dependencies, blockers, and integration risks to prevent delays. It is explicitly not just a larger daily standup for individual team status reports, which is a common misconception.

   Read the full bite: What is a Scrum of Scrums and what's shared there?

9. Question 9 of 30

   Which scenario best illustrates the purpose and appropriate content of a Scrum of Scrums?

   • aScrum Masters meet daily to compare team commitments and escalate delays to senior management for intervention.
   • bEngineering managers review and approve all cross-team architectural decisions before teams begin implementation.
   • cA rotating engineer highlights an upcoming API change, identifies affected teams, and coordinates a joint integration test to manage the rollout.
   • dEach team lead reports their team's completed stories, current velocity, and sprint burndown to the release manager.
   Show the answer

   Answer: c · A rotating engineer highlights an upcoming API change, identifies affected teams, and coordinates a joint integration test to manage the rollout.

   A Scrum of Scrums is a coordination forum where technical ambassadors surface dependencies and integration risks to resolve impediments across teams. It is not a status meeting where leads report progress or velocity metrics to management, which is a common anti-pattern.

   Read the full bite: What is a Scrum of Scrums purpose and what technical info is shared?

10. Question 10 of 30

    When multiple Scrum teams work on one product, what is the primary purpose of establishing a shared Definition of Done?

    • aTo allow each team to define its own independent quality standards, promoting team autonomy and speed.
    • bTo ensure the work from all teams integrates into a single, usable, and potentially releasable Increment each Sprint.
    • cTo provide a standardized checklist of development tasks, like code reviews and unit tests, for all teams.
    • dTo prepare all individual team outputs for a separate 'integration sprint' that occurs after several development sprints.
    Show the answer

    Answer: b · To ensure the work from all teams integrates into a single, usable, and potentially releasable Increment each Sprint.

    The correct answer is C because a shared DoD's main purpose is to ensure all work combines into a single, integrated, and usable Increment each Sprint. Option A is a common misconception; while teams have autonomy, a shared DoD is necessary to create a common quality baseline for the integrated product.

    Read the full bite: What is the purpose of a shared 'Definition of Done'?

11. Question 11 of 30

    When multiple Scrum Teams contribute to a single product, what is the primary purpose of a shared Definition of Done?

    • aTo provide a detailed checklist for each team to verify their individual tasks are completed before the Sprint Review.
    • bTo establish a consistent quality standard and ensure all teams' work integrates into a single, usable, and potentially releasable product Increment.
    • cTo define the criteria for the Quality Assurance team to perform final validation before the product is released to end-users.
    • dTo outline the specific functional and non-functional requirements for each Product Backlog Item, ensuring user satisfaction.
    Show the answer

    Answer: b · To establish a consistent quality standard and ensure all teams' work integrates into a single, usable, and potentially releasable product Increment.

    A shared Definition of Done ensures a consistent quality standard across all teams and guarantees their work integrates into a unified, usable product Increment. Option A is incorrect because it misses the crucial integration aspect for multiple teams, while Option D describes Acceptance Criteria. Option C is an anti-pattern, as the Scrum Teams are collectively responsible for the DoD, not a separate QA team.

    Read the full bite: Purpose of a Shared Definition of Done for Multiple Teams

12. Question 12 of 30

    Which scenario best illustrates when an organization should introduce ResearchOps?

    • aA scaling company where researchers spend most of their time on recruitment and consent, insights are scattered, and processes are rebuilt for each study.
    • bAn early-stage startup with one researcher who runs infrequent studies and manages their own scheduling.
    • cA team that has not yet defined its research priorities and needs help deciding what questions to answer.
    • dA mature organization that wants to replace generalist researchers with operations specialists to cut costs.
    Show the answer

    Answer: a · A scaling company where researchers spend most of their time on recruitment and consent, insights are scattered, and processes are rebuilt for each study.

    The card states ResearchOps becomes essential when research scales and logistics consume researcher time, insights are siloed, or processes are repeatedly rebuilt. Option A captures these exact scaling pain points, whereas Option B describes an early-stage context where the overhead of formal ResearchOps likely outweighs its benefits.

    Read the full bite: ResearchOps: The Pit Crew for User Research

13. Question 13 of 30

    A design team with no dedicated researchers needs fast, directional feedback on a new onboarding flow. Which approach best exemplifies research democratization?

    • aA designer runs lightweight usability sessions using a template after brief coaching, with findings later reviewed by a professional researcher
    • bAny interested team member interviews users without training because the study is informal and the stakes are low
    • cThe product manager independently designs a quantitative experiment to measure signup rates across multiple variants
    • dThe team waits until a researcher can be hired to conduct the study to avoid any risk of bias
    Show the answer

    Answer: a · A designer runs lightweight usability sessions using a template after brief coaching, with findings later reviewed by a professional researcher

    Democratization multiplies researcher impact by letting trained non-researchers run simple studies with templates and coaching, while professionals review findings and own complex work. Option B is tempting because the stakes are low, but untrained execution without oversight is explicitly warned against as a free-for-all that erodes trust, and option C wrongly assigns complex experimental design to a non-researcher.

    Read the full bite: Research Democratization: Scale Without Diluting Quality

14. Question 14 of 30

    Why is tensor parallelism typically kept within a single node while pipeline parallelism spans nodes?

    • aPipeline parallelism cannot be combined with tensor parallelism
    • bTensor parallelism all-reduces every layer needing fast interconnect; pipeline parallelism only passes activations at stage boundaries
    • cPipeline parallelism requires more bandwidth than tensor parallelism per layer
    • dTensor parallelism replicates the whole model so it must stay local
    Show the answer

    Answer: b · Tensor parallelism all-reduces every layer needing fast interconnect; pipeline parallelism only passes activations at stage boundaries

    Tensor parallelism's frequent high-volume all-reduces demand NVLink-class links, so it stays intra-node; pipeline parallelism's lighter boundary communication tolerates slower cross-node links. Neither replicates the full model, and they are commonly combined.

    Read the full bite: Tensor versus pipeline parallelism for large models

15. Question 15 of 30

    A research director finds that skilled researchers produce excellent insights, yet recruitment, governance, and tooling are inconsistent across divisions. What should the ResearchOps Maturity Matrix diagnose?

    • aWhich divisions need additional research training to manage recruitment and privacy more effectively.
    • bWhether individual study quality and research methods are rigorous enough to support scaling.
    • cThe degree to which UX maturity and research adoption vary across the organization.
    • dGaps in the operational infrastructure governing participant pipelines, data protocols, and tooling strategy.
    Show the answer

    Answer: d · Gaps in the operational infrastructure governing participant pipelines, data protocols, and tooling strategy.

    The matrix audits operational infrastructure such as participant pipelines and governance, not the quality of insights or researcher skill. Although inconsistent recruitment and privacy practices might suggest a training gap, the matrix reveals that fragmented operational systems—not skill deficiencies—are the true bottleneck.

    Read the full bite: ResearchOps Maturity Matrix: Built for Operations

16. Question 16 of 30

    From an engineer's perspective, how do SAFe and LeSS primarily differ regarding planning and team autonomy?

    • aSAFe involves engineers in highly structured, multi-day Program Increment (PI) planning with less team autonomy, while LeSS promotes decentralized, team-led sprint planning and maximizes team autonomy.
    • bLeSS prioritizes explicit dependency management via a program board, whereas SAFe empowers teams with full control over their feature delivery.
    • cSAFe offers greater team autonomy through continuous planning, while LeSS centralizes planning in large, multi-day events.
    • dBoth frameworks aim for maximum team autonomy, but SAFe uses a Release Train Engineer for planning, and LeSS relies on a System Architect.
    Show the answer

    Answer: a · SAFe involves engineers in highly structured, multi-day Program Increment (PI) planning with less team autonomy, while LeSS promotes decentralized, team-led sprint planning and maximizes team autonomy.

    SAFe is characterized by its highly structured, top-down Program Increment (PI) planning events, which inherently lead to less team autonomy. In contrast, LeSS promotes decentralized, team-led sprint planning and is designed to maximize team autonomy. Option C is incorrect because it reverses these core characteristics for both frameworks.

    Read the full bite: SAFe vs. LeSS: Planning, Dependencies, and Autonomy

17. Question 17 of 30

    Which application scenario is least suitable for relying on a read replica for data retrieval?

    • aA user verifying a newly updated password immediately after the change.
    • bA website displaying a product catalog to many concurrent users.
    • cA public-facing blog where new articles are published and then read by many.
    • dA business intelligence tool generating daily sales reports.
    Show the answer

    Answer: a · A user verifying a newly updated password immediately after the change.

    The card states that read replicas are unsuitable when immediate read-after-write consistency is required, such as verifying a newly changed password, due to potential replica lag. The other scenarios involve read-heavy workloads where some data staleness is acceptable.

    Read the full bite: Read Replicas: Scale Out Your Database Reads

18. Question 18 of 30

    An organization wants to scale agile while maximizing team autonomy and minimizing formal, top-down planning ceremonies. Which approach should they adopt?

    • aSAFe, because its structure of smaller Agile Release Trains inherently reduces the need for cross-team coordination.
    • bSAFe, because its Program Board provides a formal, visible way to manage all cross-team dependencies during PI Planning.
    • cLeSS, because it favors direct team-to-team communication and shared code ownership to resolve dependencies as they arise.
    • dLeSS, because the Product Owner is responsible for deconflicting all dependencies before Sprint Planning begins.
    Show the answer

    Answer: c · LeSS, because it favors direct team-to-team communication and shared code ownership to resolve dependencies as they arise.

    LeSS is designed to maximize team autonomy by encouraging teams to self-organize and resolve dependencies directly, contrasting with SAFe's emphasis on formal, large-scale planning events like PI Planning to ensure alignment.

    Read the full bite: SAFe vs. LeSS: Planning, Dependencies, and Autonomy

19. Question 19 of 30

    What is the primary purpose of organizing multiple Agile teams into an Agile Release Train (ART)?

    • aTo enable rapid formation and disbandment of teams for temporary projects.
    • bTo ensure synchronized delivery of integrated value across multiple teams working on a complex solution.
    • cTo centralize decision-making and reduce overhead for small, single-team products.
    • dTo allow each team complete autonomy and independent release schedules for maximum flexibility.
    Show the answer

    Answer: b · To ensure synchronized delivery of integrated value across multiple teams working on a complex solution.

    An ART is designed to align and synchronize multiple teams working on a large, complex solution, ensuring their integrated efforts deliver value together. It is explicitly not for temporary projects, small products, or promoting independent release schedules, as that would defeat the purpose of synchronization.

    Read the full bite: Release Trains: Aligning Multiple Agile Teams

20. Question 20 of 30

    How does LeSS fundamentally change a manager's approach to cross-team dependencies compared with traditional program management?

    • aLeSS replaces command meetings with team self-coordination and manager-as-teacher enablement
    • bLeSS introduces additional project managers to coordinate dependencies across teams
    • cLeSS scales Scrum by adding coordination meetings while keeping the same management hierarchy
    • dLeSS managers use detailed plans and Gantt charts to force alignment by command
    Show the answer

    Answer: a · LeSS replaces command meetings with team self-coordination and manager-as-teacher enablement

    LeSS shifts from command to enablement, relying on team self-coordination and managers as teachers who remove systemic barriers. Option B is tempting because scaling often suggests adding coordinators, but LeSS explicitly rejects inserting more project managers to handle dependencies.

    Read the full bite: How does management evolve in scaled agile versus traditional program management?

21. Question 21 of 30

    How does the cluster module improve a Node web server's throughput on a multi-core machine?

    • aIt increases the size of the single process's heap
    • bIt makes each individual request execute on multiple cores simultaneously
    • cIt converts the event loop into a multi-threaded scheduler
    • dIt forks multiple worker processes sharing one port so requests spread across cores
    Show the answer

    Answer: d · It forks multiple worker processes sharing one port so requests spread across cores

    cluster runs several worker processes on a shared listening socket so connections are distributed across cores. A single request still runs within one worker's event loop; cluster does not parallelize one request or multithread the loop.

    Read the full bite: Purpose of the Node.js cluster module

22. Question 22 of 30

    In a federated design system governance model, how are responsibilities typically divided between the contributing product team and the core team?

    • aProduct teams build and maintain components; the core team sets standards and gatekeeps merges
    • bThe core team builds everything while product teams only file requests
    • cEach team forks the system and maintains its own independent copy
    • dProduct teams own the design tokens while the core team writes the components
    Show the answer

    Answer: a · Product teams build and maintain components; the core team sets standards and gatekeeps merges

    Federation distributes building and maintenance to product teams while the core team stewards standards, tokens, and the merge gate to keep consistency. Centralized building, handing tokens to teams, or forking the system all defeat the model's purpose.

    Read the full bite: A federated design system governance model

23. Question 23 of 30

    Which workload is the textbook fit for the cluster module rather than worker_threads?

    • aResizing a large uploaded image off the event loop
    • bComputing an expensive Argon2 password hash without blocking
    • cScaling an IO-bound HTTP API to use all CPU cores with crash isolation
    • dSharing a large buffer cheaply between two units of work
    Show the answer

    Answer: c · Scaling an IO-bound HTTP API to use all CPU cores with crash isolation

    cluster forks independent processes sharing a port to scale IO-bound request throughput across cores with isolation. The image-resize, hashing, and shared-buffer cases are CPU-bound or shared-memory tasks that suit worker_threads.

    Read the full bite: worker_threads versus cluster: when to use each

24. Question 24 of 30

    What is a fundamental requirement for an application to effectively leverage horizontal scaling?

    • aIt must be capable of processing requests at a significantly higher speed than a single server.
    • bIt must ensure that all critical data is stored directly on each individual server instance.
    • cIt must be designed to be stateless, allowing any instance to handle any user request.
    • dIt must use a specialized, high-performance database that can only run on a single machine.
    Show the answer

    Answer: c · It must be designed to be stateless, allowing any instance to handle any user request.

    Horizontal scaling requires applications to be stateless so that any instance can handle any request, with shared state managed externally. Option B describes a stateful application, which is explicitly stated as a 'footgun' and a reason *not* to use horizontal scaling.

    Read the full bite: Horizontal Scaling: Add More Machines, Not Bigger Ones

25. Question 25 of 30

    In data-parallel distributed training, what is the main reason adding more GPUs rarely gives perfectly linear speedup?

    • aThe model gets larger as GPUs are added
    • bEach GPU trains on the entire dataset independently
    • cMixed precision must be disabled when scaling out
    • dGradient synchronization across GPUs adds communication overhead
    Show the answer

    Answer: d · Gradient synchronization across GPUs adds communication overhead

    Data parallelism must all-reduce gradients every step, and that network communication grows with GPU count, capping speedup. Each GPU processes only a shard of the batch, mixed precision still applies, and the model size is unchanged.

    Read the full bite: How would you speed up slow single-GPU training?

26. Question 26 of 30

    A team has drafted a comprehensive ResearchOps playbook. What is the most critical next step to ensure it succeeds as a living process?

    • aTest it on a real project and schedule regular updates
    • bExpand it to cover every possible research method before any rollout
    • cLock the final version and require strict adherence across all squads
    • dStandardize outputs but let researchers keep files in personal drives
    Show the answer

    Answer: a · Test it on a real project and schedule regular updates

    The card warns that the biggest footgun is writing a playbook once and letting it rot, so testing it on a real project and scheduling regular updates is essential. Expanding it indefinitely (B) delays value, while locking it (C) creates the exact rigid, outdated document the playbook is meant to avoid.

    Read the full bite: ResearchOps Playbooks: Standardize UX Research

27. Question 27 of 30

    You have two Node.js servers behind a load balancer. Client A connects to server 1, client B to server 2. Server 1 calls io.emit('update', data). Which clients see the update?

    • aNeither client, because they are on different servers.
    • bOnly client A on server 1, because the default adapter is local only.
    • cBoth A and B, because io.emit() is global.
    • dOnly B on server 2, because the load balancer routes the event there.
    Show the answer

    Answer: b · Only client A on server 1, because the default adapter is local only.

    The default in-memory adapter only broadcasts to sockets on the same server instance. Client A sees the update, but B doesn't. Using a Redis adapter would make both see it. The load balancer routes client connections, not io.emit() messages; events don't travel through the load balancer.

    Read the full bite: Cross-server Socket.IO communication in horizontal scaling?

28. Question 28 of 30

    What is the primary risk that a federated design system contribution model must actively guard against?

    • aSlower contribution throughput than a centralized team
    • bInconsistency and quality drift without strong governance and review
    • cInability to publish packages to a registry
    • dA single point of failure if the core team is unavailable
    Show the answer

    Answer: b · Inconsistency and quality drift without strong governance and review

    Federation distributes contribution, so the main hazard is fragmentation and quality drift, which governance and review counter. Slow throughput and single points of failure are characteristic of centralized models, not federated ones.

    Read the full bite: Centralized vs federated design system team models

29. Question 29 of 30

    What is the primary problem SAFe (Scaled Agile Framework) aims to solve for organizations?

    • aProviding a lightweight, adaptable framework for small, co-located development groups.
    • bStandardizing project management practices across an enterprise to reduce all overhead.
    • cEnhancing the efficiency and autonomy of individual Scrum teams.
    • dCoordinating the efforts of many interdependent teams working on a large, complex product.
    Show the answer

    Answer: d · Coordinating the efforts of many interdependent teams working on a large, complex product.

    SAFe is designed to apply Agile principles across large organizations, specifically to coordinate multiple teams (50+ people) working on a single, complex product. While it involves standardization, its main goal is not to reduce all overhead, and it is explicitly stated as overkill for small groups or single teams.

    Read the full bite: SAFe: Scaling Agile Beyond a Single Team

30. Question 30 of 30

    What is the fundamental approach LeSS uses to scale Scrum?

    • aEmpowering each team to have its own Product Owner and Definition of Done.
    • bDescaling the organization by applying single-team Scrum principles to multiple teams on one product.
    • cIntroducing a hierarchy of Product Owners to manage multiple product backlogs.
    • dStandardizing processes across independent teams working on separate product components.
    Show the answer

    Answer: b · Descaling the organization by applying single-team Scrum principles to multiple teams on one product.

    LeSS scales Scrum by simplifying the organizational structure and applying one-team Scrum principles to multiple teams working on a single product. It explicitly avoids introducing multiple Product Owners or separate backlogs, which is why option C is incorrect.

    Read the full bite: Large-Scale Scrum (LeSS): Scaling Scrum by Descaling the Org