Observability Data Engineer Interview Questions

Option B is strongest because it gives a precise definition — total unique time series from label combinations — provides a concrete example with user_id, explains the memory mechanism behind the problem, and gives a practical mitigation at the OpenTelemetry collector layer. Option A defines cardinality at a label level rather than at the time series level, which is the definition that matters for metrics backends. Option C makes the excellent architectural point about reserving high-cardinality identifiers for traces and logs rather than metrics, which is the right design principle, but it does not explain the memory mechanism. Option D introduces the cross-product danger and detection tooling — tsdb analysis and per-label limits — which are production-grade practices, but the cross-product explanation, while accurate, is harder to follow quickly in an interview. Observability interview best practice: explain cardinality at the time series level, not just the label level, and name one specific mitigation you apply at the collection layer.

Option B is strongest because it explains the three-stage architecture — receivers, processors, exporters — with protocol examples, and then articulates three concrete advantages over direct instrumentation. The tail-based sampling and cardinality reduction points show deep observability domain knowledge. Option A is correct but purely descriptive; it does not explain the architecture or the advantages over direct instrumentation. Option C identifies the decoupling benefit correctly and gives a realistic example, but it covers only one of the three advantages and does not explain the processor stage. Option D describes the two-tier gateway topology, which is a mature production architecture, and the tail sampling rationale — full trace assembly — is excellent, but it does not explain the basic architecture for an interviewer who may be unfamiliar with the collector. Observability interview best practice: name all three pipeline stages — receivers, processors, exporters — before describing the advantages; the processor stage is what distinguishes the collector from a simple proxy.

Option B is strongest because it defines both types precisely, identifies the fundamental limitation of head-based sampling — blindness to outcome — and explains what tail-based sampling enables: retaining 100% of error and slow traces while aggressively sampling healthy ones. It also names the architectural requirement: a buffering gateway. Option A is a correct one-sentence definition but lacks any analysis of the trade-offs. Option C makes the important stateless versus stateful distinction and explains why tail sampling must be centralised, which is a key architectural insight, but it does not describe the decision criteria that make tail sampling valuable — keeping errors and high-latency traces. Option D describes a hybrid strategy, which is the most practical production approach, and names the specific OTel processor, but it does not define the two sampling types clearly enough for an interviewer who may not know them. Observability interview best practice: describe the limitation of head-based sampling first — cannot know outcome at decision time — then explain how tail-based sampling removes that limitation.

Option B is strongest because it names all three layers with specific tool examples, explains the rationale for each — local buffering for outage resilience, message queue for spike absorption — and ends with the key reliability property: the pipeline survives indexer slowdowns without dropping logs. Option A describes the basic topology without addressing backpressure or reliability. Option C makes the two key structural points — structured logs and message queue — with clear reasoning, but it omits the local agent buffering layer, which is what prevents data loss during network interruptions. Option D focuses on dimensioning and testing, which is excellent operational thinking, and the per-category retention policy is a mature cost-management practice, but it skips the pipeline architecture that the question asks about. Observability interview best practice: name all three pipeline layers and explain the specific failure mode each one prevents; this shows you have designed for reliability, not just functionality.

Option B is strongest because it identifies the root cause — ad-hoc metric additions without governance — and addresses it with three complementary practices: catalogue, CI validation, and cardinality budgets. The quarterly deprecation audit shows long-term pipeline health thinking. Option A describes a documentation process without enforcement; documentation alone does not prevent drift. Option C makes the excellent semantic versioning point for metrics, which is underused in practice, and the stable alias pattern is a clever mitigation, but it focuses only on the accuracy problem and not on the governance and cost dimensions. Option D treats metrics as an API contract, which is a mature framing, and the change request process with blast radius assessment is production-grade governance, but it focuses on the change management process without describing ongoing monitoring of metric health. Observability interview best practice: combine a governance mechanism — catalogue and CI validation — with a cost mechanism — cardinality budgets — to show you think about both correctness and operational sustainability.