1. Clarify Requirements First

Appears in: Telemetry Ingestion Pipeline — this is §1 of the full design, split into its own file so the root stays a table of contents.

1. Clarify Requirements First

The first 5 minutes of the interview is yours to steer. Ask these; the answers change the entire architecture.

Signal types

• Metrics only, or metrics + logs + traces (unified MELT pipeline)?
• Are traces sampled at head or tail?

  This determines whether spans need to be held in memory for assembly.

• Any profiling signals (continuous profiling, eBPF)?

Scale envelope

Dimension	Small (startup)	Mid (Netflix-class)	Large (hyper-scale)
Agent count	~1K	~100K	~10M+
Metric series (active)	~10M	~1B	~100B+
Ingest rate	~1M pts/sec	~500M pts/sec	~50B pts/sec
Trace spans/sec	~100K	~10M	~1B
Log lines/sec	~1M	~500M	~50B

At MAANG interviews, assume Netflix/Google scale unless told otherwise.

Consistency and durability

• Can we drop data during a rolling restart? (Usually: yes for metrics, no for billing traces.)
• What is the maximum acceptable ingestion lag before data lands queryable? (SLO: typically < 60s for metrics, < 5 min for traces.)
• Retry policy: at-least-once or exactly-once? Exactly-once is expensive — push back unless there is a billing requirement.

Multi-tenancy

• Single tenant (internal platform) or multi-tenant SaaS?
• Per-tenant quota enforcement? Isolation at the storage layer or earlier?

Protocol

• Must support OTLP? Prometheus remote-write? Both? Datadog agent wire format?
• Is the interviewer implying a greenfield (pick the best protocol) or a brownfield (must accept legacy agents)?