System design practice · September 23, 2026 evidence note

This review guided the twelve constructed problems distributed across system design, observability, reliability, data at scale, AI systems, and migrations. It compares publicly reported interviews with engineering accounts of real failure and scale. Interview anecdotes are self-reported, sometimes lack the original interview date, and cannot establish a company's official question bank or the frequency of any prompt. New exercises are original scenarios with invented workloads and explicit contracts.

Recent interview signals

Public report	Published	Specific signal used	Limit
LinkedIn candidate report	January 21, 2026	An API quota/rate-tracker system-design prompt is described	One anonymous report; details and outcome cannot be verified
2026 ExperiencedDevs discussion	July 8, 2026	Respondents discuss coding, design, practical implementation, and AI tooling	Commenters have different roles and companies
Senior full-stack interview report	January 5, 2026	Multiple rounds include technical coding and system-design discussion	Anonymous and self-reported; no universal round format follows
Senior system design preparation discussion	September 2026	Candidates want a way to distinguish mid-level from senior reasoning and test their responses	Represents a preparation gap, not an interview question count

Recent engineering cases that shaped the follow-ups

Primary account	Date	Practice decision prompted by it
Meta: modernizing Groups search	April 21, 2026	Exercise hybrid retrieval, result eligibility, and measurable relevance
Spotify: separate personalization and experimentation stacks	January 7, 2026	Separate ranking latency and production safety from evaluation
Spotify: engineering behind 2025 Wrapped Archive	March 12, 2026	Test launch capacity, replay, recovery, and price under a spike
Meta: instant power loss readiness	June 3, 2026	Make regional failure abrupt in the exercise and verify recovery rather than assuming it

Current service behavior was checked separately against AWS documentation: standard queue delivery, FIFO send deduplication, DynamoDB read consistency, and DynamoDB global table modes. These technical references are not dated interview evidence.

What the new exercise set covers

After the lesson	New prompts	Distinct failure tested
System design method	Quotas, ticket inventory, chat, feed, video, checkout	Concurrent authority, replay, hot fanout, byte paths, external side effects
Observability	Slow request	Average latency hides wait and retry amplification
Reliability	Durable jobs	Queue acknowledgement differs from completed work
Data at scale	Document search, trending counts	ACL revocation and hot event-time partitions
AI systems	Personalized ranking	Eligibility, deadline budgets, experimental evidence
Migrations	Regional failover	Acknowledged writes missing after asynchronous replication

The exercises use explicit inputs, a candidate opening, a before/after diagram, a failure timeline, and senior/staff follow-ups. Eight also carry a third focused visual for a capacity calculation or disputed state. No scenario is attributed to a company unless a source supports that narrow attribution; the exercise itself is always labeled as constructed.

Second set: service boundaries, September 23, 2026

Nine more constructed prompts were placed after their prerequisites: tenant isolation and audit after security; collaborative editing and webhooks after system-design method; feature rollout after delivery; event ingestion after streaming and data-at-scale; overload after performance and cost; knowledge assistant after AI evaluation; erasure after migration and recovery. Each has a failure/fix visual, an AWS architecture with both the AWS product and its general role on the box, a concrete four-state trace, input-to-outcome cases, and alternative service choices. The earlier twelve design prompts also acquired AWS service/role diagrams and a rationale for key substitutions.

Recently published engineering evidence	Date	Decision translated into an original exercise	What it does not establish
Meta: ingestion system migration	May 12, 2026	Replayable ingestion, coexistence and measured rollback	The exact event-ingestion or feature-flag interview prompt
Meta: privacy-aware asset classification	June 25, 2026	Trace copies, ownership, and erasure exceptions across derived data	The legal retention rule in any jurisdiction
Meta: ZGateway proxy	September 3, 2026	Admission, isolation and measured overload protection	An official interview question or a universal capacity number
Spotify: context layer for data assistant	June 10, 2026	Named ownership, trusted retrieval context and limits of generated answers	Spotify's assistant system design interview rubric

Current AWS behavior was checked separately with official documentation: AppConfig rollback and alarms, EventBridge HTTP destination timeouts/retries, Kinesis hot-key scaling, DynamoDB pooled tenant isolation, Bedrock Knowledge Bases sync, and S3 Object Lock. Those docs describe service mechanisms; they are not dated interview reports. A July 2026 ExperiencedDevs discussion signals continuing use of design and practical integration but does not establish frequency of any of these prompts. We chose the problems for distinct architectural decisions, not for an invented "most asked" ranking.

Sources and further reading · 22