Scores are estimates, not hiring predictions.
17 Sep 2026, 01:48 · Backend engineering · Senior · General calibrated coverage (IST)
Published · revision 1
Strong evidence-led diagnosis. The next improvement is to make trade-off boundaries and rejected alternatives more explicit.
88% evidence confidence
Scoring rubric: technical-core-v1
“I would first establish request shape, tenant fairness, burst size and failure tolerance. I would compare a token bucket at the gateway with per-tenan…”turns[0].answer · SHA-256 1cdbd8355e40
“I would capture the slow query and execution plan before changing indexes. I would compare row estimates with actuals, inspect buffer reads, lock wait…”turns[1].answer · SHA-256 4a6f0b5c5e11
“I would stop or roll back the rollout using the deployment marker rather than debating dashboards. I would preserve logs and traces, compare affected …”turns[2].answer · SHA-256 380e90957cea
The answer establishes evidence before action and connects the decision to an observable outcome.
I would first establish request shape, tenant fairness, burst size and failure tolerance. I would compare a token bucket at the gateway with per-tenant counters in Redis, choose the simplest policy that preserves fairness, and validate it using rejected-request rate, queue time and downstream saturation.
The answer establishes evidence before action and connects the decision to an observable outcome.
I would capture the slow query and execution plan before changing indexes. I would compare row estimates with actuals, inspect buffer reads, lock waits and table statistics, then reproduce against representative volume. I would test the smallest index or query change and verify p95 latency and write amplification.
The answer establishes evidence before action and connects the decision to an observable outcome.
I would stop or roll back the rollout using the deployment marker rather than debating dashboards. I would preserve logs and traces, compare affected cohorts and verify the rollback at the user-facing SLI. Once stable, I would reconcile metric definitions and record the missing alert or release guard.
Confirmed transcript
Evidence for result revision 1.
Correct transcription only. Saves retain earlier versions; material changes request a new evaluation.
I would first establish request shape, tenant fairness, burst size and failure tolerance. I would compare a token bucket at the gateway with per-tenant counters in Redis, choose the simplest policy that preserves fairness, and validate it using rejected-request rate, queue time and downstream saturation.
Correct transcription errors, not your original answer.
I would capture the slow query and execution plan before changing indexes. I would compare row estimates with actuals, inspect buffer reads, lock waits and table statistics, then reproduce against representative volume. I would test the smallest index or query change and verify p95 latency and write amplification.
Correct transcription errors, not your original answer.
I would stop or roll back the rollout using the deployment marker rather than debating dashboards. I would preserve logs and traces, compare affected cohorts and verify the rollback at the user-facing SLI. Once stable, I would reconcile metric definitions and record the missing alert or release guard.
Correct transcription errors, not your original answer.