Intermittent failures are non-determinism: timing, ordering, or resource state. The interviewer wants to hear that you attack non-determinism systematically instead of guessing. The discipline is: reproduce, instrument, bisect, then fix — in that order.
Reproduce first, because "intermittent" means the reproduction is probabilistic. Keep every failed request's trace ID and log lines; correlate the failure window with deploys, traffic, and GC pauses. If you can't reproduce in prod, replay the captured requests against a shadow environment — same payloads, same ordering, same timing. Chaos is the honest reproduction tool: kill a node, throttle a link, induce the partition, and watch whether the failure appears.
Instrument before hypothesizing. You need three signals correlated by request ID, not by timestamp: metrics (p50/p99 latency, error rates), structured logs (every retry, every timeout, every backoff), and distributed traces across the call graph. Correlating by timestamp is itself a trap — NTP skew between machines can exceed the margin you're trying to measure; monotonic clocks and trace IDs are the only trustworthy ordering.
Then bisect the space: which layer (client, LB, service, DB, network), which tail (the p99 is where intermittency lives — average latency hides it), and which condition. The most common root causes have boring names: a timeout set too tight against a GC pause or cold start; a connection pool exhausted by a traffic spike; replica lag serving stale reads; a hot key overloading one partition; retry storms turning one slow node into a cascading failure. The classic amplification: each failed request retries with backoff, the retries land on the same busy node, timeouts compound, and a 1% failure rate becomes a regional incident.
The fix is usually to make the failure visible and bounded rather than to chase the specific bug: circuit breakers so failures fail fast, retries with jittered exponential backoff, idempotency keys so retries are safe, and hedging for tail latency. Say the closing line: "an intermittent bug I can't reproduce is a system I haven't instrumented yet — I change the system until failures become deterministic enough to observe."