API-001 – OPS-001

Backend, MLOps, and safety

0 of 25 done · 0 due for review.

API-001Design the inference request/response contract: input text, output label and confidence, plus a model version.construct · a-model-needs-a-contractnew
API-002Separate client errors from server errors: missing 'text' vs model timeout. Which status class does each get, and why does it matter?compare · a-model-needs-a-contractnew
API-003Design health and readiness checks: the model loads for 20 seconds, then can serve. What does each check report and when?construct · serve-the-model-safelynew
API-004Version a model endpoint safely: v1 is live, v2 is ready. Design the rollout that keeps both reachable.construct · serve-the-model-safelynew
API-005Validate an oversized request: 50 MB of text hits a 1 KB limit. Respond correctly and keep the service alive.operate · serve-the-model-safelynew
QUEUE-001Choose synchronous versus asynchronous execution: a search that takes 3 seconds vs an ingestion job that takes 3 minutes.predict · timeouts-and-retries-are-not-freenew
QUEUE-002Make this job retry idempotent: a worker died after committing row 400 of 1000. Write the retry rule.implement · timeouts-and-retries-are-not-freenew
QUEUE-003Handle this dead-letter record: a row failed validation 5 times. What do you do with it, and what must the record carry?operate · timeouts-and-retries-are-not-freenew
QUEUE-004Design backpressure for a busy worker: the queue grows 5× faster than workers drain it.construct · timeouts-and-retries-are-not-freenew
QUEUE-005Recover this worker after its lease expires: the worker processed a job but never renewed the lease.operate · timeouts-and-retries-are-not-freenew
PERF-001Choose a batching boundary: 1000 tiny requests at 5 ms each vs one batch at 400 ms. Which wins at high load, and why?predict · measure-the-service-not-just-the-modelnew
PERF-002Design a cache key for inference: identical texts should hit; texts that differ must not.construct · measure-the-service-not-just-the-modelnew
PERF-003Diagnose this p99 latency spike: p50 is 8 ms, p99 is 900 ms, once per minute.debug · measure-the-service-not-just-the-modelnew
PERF-004Estimate cost per request: GPU at $1/hour serves 3600 requests/hour. What is the per-request compute cost?derive · measure-the-service-not-just-the-modelnew
PERF-005Choose CPU or GPU for this workload: 10k texts/day, 200 ms CPU latency is fine, batch of 32.predict · measure-the-service-not-just-the-modelnew
MLOPS-001Compare two model artifacts by hash and metrics: artifact A (sha 3f9a, F1 0.82) vs artifact B (sha 3f9a, F1 0.85).compare · reproducibility-is-part-of-correctnessnew
MLOPS-002Design a promotion gate: a candidate model must pass three checks before production. List them.construct · measure-the-service-not-just-the-modelnew
MLOPS-003Roll back this bad model version: v7 was promoted an hour ago; quality dropped 20%. What is the rollback sequence?operate · measure-the-service-not-just-the-modelnew
MLOPS-004Choose logs, metrics, and traces for this failure: p99 latency doubled after deploy X. Which instrument answers which question?compare · measure-the-service-not-just-the-modelnew
MLOPS-005Detect data drift before quality collapses: live traffic now uses 40% new vocabulary, model F1 still looks fine. Which alert fires first?debug · find-the-error-familynew
SAFE-001Identify PII in this ticket: 'contact ana@example.com at +1-555-0100 for the invoice'. Design the redaction.implement · serve-the-model-safelynew
SAFE-002Design retention and deletion behavior: tickets must be deleted 30 days after closure. What does deletion actually mean for the model?construct · serve-the-model-safelynew
SAFE-003Build a prompt-injection fixture: a document says 'ignore the system prompt'. Write the test case and the expected safe output.implement · cite-the-evidencenew
SAFE-004Find an unsafe tool boundary: a search tool can read internal tickets, and the model freely calls it. What boundary is missing?debug · cite-the-evidencenew
OPS-001Write the incident timeline from these traces: 14:02 latency rises, 14:05 quality drops, 14:07 deploy rolled back, 14:11 recovered. Write the postmortem skeleton.communicate · measure-the-service-not-just-the-modelnew

Opening…