Drill bank
Sixty-six questions from the source document. Ninety seconds each, out loud, no notes. If you cannot, that is your study list, and this page is how you find the page that teaches it.
Answers are written in spoken form: what you would say in a room, not what you would write in a design doc. That is the distinction the source document's answer key was missing. A written answer can be dense and assume the reader will re-read a sentence; a spoken answer needs a shape the listener can follow in real time, which means leading with the conclusion, keeping to three points, and stopping.
Written batches
- Drills 1 to 6: AI, LLM and RAG
- Drills 10 to 15: streaming, reactive and APIs
- Drills 37 to 42: leadership
The full map
Every drill, with the page that teaches it. All 66 now resolve to a full
treatment. Where a drill also appears in a written batch above, the spoken link
gives the ninety-second version you would actually say in a room; the other link is
the page that teaches the material behind it.
AI and LLM (1 to 6)
| # | Drill | Where it is taught |
|---|---|---|
| 1 | Why is prefill compute-bound and decode memory-bandwidth-bound? | Prefill vs decode · spoken |
| 2 | Your RAG answers are confidently wrong. Walk the diagnosis. | Confidently wrong RAG · spoken |
| 3 | When would you fine-tune instead of improving retrieval? | spoken |
| 4 | How do you evaluate an agent that takes 20 steps? | Evaluating an agent · spoken |
| 5 | How do you defend against indirect prompt injection? | Prompt injection · spoken |
| 6 | Cut LLM spend 60 percent. What is the order of moves? | Cost engineering · spoken |
Context and graph engineering (7 to 9)
| # | Drill | Where it is taught |
|---|---|---|
| 7 | Your prompt is 3,000 tokens of business rules. What do you do? | Budgeting a context window |
| 8 | When does a knowledge graph beat a vector index? | Knowledge graph vs vector |
| 9 | How do you budget a 128k context window for an agent? | Budgeting a context window |
Streaming, reactive and APIs (10 to 15)
| # | Drill | Where it is taught |
|---|---|---|
| 10 | Explain exactly-once in Kafka and what it does not cover. | Kafka exactly-once · spoken |
| 11 | flatMap vs concatMap: behaviour, and when each is wrong. | flatMap vs concatMap · spoken |
| 12 | Do virtual threads make WebFlux obsolete? | Virtual threads vs reactive · spoken |
| 13 | How do you fix N+1 in GraphQL, and why does caching not solve it? | GraphQL N+1 · spoken |
| 14 | Why does an L4 load balancer break gRPC? | L4 and gRPC · spoken |
| 15 | Watermarks in Flink: what problem do they solve? | Watermarks · spoken |
Caching and edge (16 to 18)
| # | Drill | Where it is taught |
|---|---|---|
| 16 | Prevent a cache stampede on a hot key. Three approaches. | Cache stampede |
| 17 | Invalidate cached content with complex dependencies. How? | Invalidation with dependencies |
| 18 | SSE or WebSocket for streaming LLM tokens, and why? | SSE vs WebSockets |
Compute, Kubernetes and OS (19 to 23)
| # | Drill | Where it is taught |
|---|---|---|
| 19 | A pod is healthy but slow. Diagnose in order. | A pod is healthy but slow |
| 20 | Why might removing CPU limits improve latency? | CPU limits and CFS throttling |
| 21 | Serverless or containers for this workload? Walk your math. | Serverless or containers |
| 22 | Where does fsync fit in a durability guarantee? | fsync and durability |
| 23 | How would you use eBPF to debug intermittent latency? | eBPF for latency |
Storage (24 to 28)
| # | Drill | Where it is taught |
|---|---|---|
| 24 | Design a DynamoDB table for these five access patterns. | DynamoDB single-table design |
| 25 | Why did our Cassandra range query start timing out? | Cassandra tombstones |
| 26 | Pick a Mongo shard key for this workload and defend it. | MongoDB shard key |
| 27 | Reindex OpenSearch with zero downtime. | Zero-downtime reindex |
| 28 | Postgres write skew under REPEATABLE READ. What happens? | Write skew |
Security (29 to 33)
| # | Drill | Where it is taught |
|---|---|---|
| 29 | Why PKCE for a confidential client? | PKCE |
| 30 | Design token revocation with stateless JWTs. | Revoking stateless JWTs |
| 31 | A critical CVE drops in a library you use. Walk the first 4 hours. | CVSS, EPSS and KEV |
| 32 | Explain zero trust to a director in 60 seconds, then the sequencing. | Zero trust |
| 33 | CVSS vs EPSS, and which drives your patching? | CVSS, EPSS and KEV |
Frontend (34 to 36)
| # | Drill | Where it is taught |
|---|---|---|
| 34 | Half our Redux store is server data. What is the migration and why? | Redux to a server cache |
| 35 | Our INP is bad. Diagnose. | Diagnosing a bad INP |
| 36 | When are micro-frontends worth it? | Micro-frontends |
Leadership (37 to 42)
| # | Drill | Where it is taught |
|---|---|---|
| 37 | Your best engineer's PR comments are demoralising juniors. First move? | The toxic code reviewer · spoken |
| 38 | Review queue depth doubled after an AI tooling rollout. What now? | spoken |
| 39 | Your director wants a date you cannot commit to. | The impossible date · spoken |
| 40 | Make the case for 25 percent reliability investment to a product VP. | Reliability investment case · spoken |
| 41 | Two teams are building the same service. You have no authority. | Two teams, same service · spoken |
| 42 | An engineer wants promotion; they are one level of scope short. | One level short · spoken |
SRE, architecture and delivery (43 to 52)
| # | Drill | Where it is taught |
|---|---|---|
| 43 | Design burn-rate alerts for a 99.9 percent SLO. Why two windows? | Burn-rate alerting |
| 44 | A label explosion took down Prometheus. What happened? | Cardinality |
| 45 | How do you test 40 microservices without a full environment? | Contract testing |
| 46 | Deploy vs release: explain the distinction and what it buys you. | Deploy is not release |
| 47 | Split this 15-person team. Walk your reasoning. | Team Topologies |
| 48 | When is a modular monolith right, and what forces an extraction? | Modular monolith vs microservices |
| 49 | Your experiment won its metric and tripped a latency guardrail. Ship? | Canary vs A/B testing |
| 50 | "We cannot lose data and cannot be down." What do you say next? | RTO and RPO |
| 51 | An untested assumption in your architecture. How would you chaos-test it? | Chaos-testing an assumption |
| 52 | You have been down-levelled in an offer. What do you do? | Being down-levelled |
Patterns and seasonality (53 to 60)
| # | Drill | Where it is taught |
|---|---|---|
| 53 | Explain the repository pattern, then argue against it here. | Repository pattern |
| 54 | "Is CQRS a good idea for us?" Walk the adoption ladder. | CQRS adoption ladder |
| 55 | Our vendor offers 99.5 percent and we sell 99.9 percent. Options? | Burn-rate alerting (dependency math) |
| 56 | A question you genuinely cannot answer. Perform the sequence. | The staff-level design interview (follow-ups) |
| 57 | 90 seconds with the CTO at the end of the loop. What do you ask? | Reverse due diligence |
| 58 | Rewrite one of your SCOR stories as STAR on the spot. | SCOR, STAR and scar tissue |
| 59 | Three red flags from the reverse-diligence list, and the questions. | Reverse due diligence |
| 60 | A 30-second scar-tissue story that fits inside a caching answer. | SCOR, STAR and scar tissue |
Deployment (61 to 66)
| # | Drill | Where it is taught |
|---|---|---|
| 61 | A PM says "the canary shows the feature is winning." Correct it. | Canary vs A/B testing |
| 62 | Canary schedule for a payments service at 200 QPS. What can you not detect? | Canary vs A/B testing |
| 63 | Shadow-test a rewrite of a service that sends emails. | Canary vs A/B testing (follow-ups) |
| 64 | A rollback fails because the old version cannot read new cache entries. | Expand and contract |
| 65 | Mobile app crashes for 2 percent of users at 40 percent rollout. | Mobile crash at 40% rollout |
| 66 | Rename a column across a 3-version compatibility window. Sequence it. | Expand and contract |
Beyond the source's 66
Questions the expansion added, because the pages exist and interviewers ask them:
| Drill | Where it is taught |
|---|---|
| Why does a service with a good p99 have a bad user-facing p50? | The tail at scale |
| How many instances do you need? | Little's Law and queueing |
| Why is exactly-once delivery impossible? | Idempotency |
| Pre-filter or post-filter a vector search, and why? | Filtered vector search |
| Why is interleaving more sensitive than an A/B test? | Interleaving and position bias |
| Why does RRF beat score normalisation? | Hybrid retrieval and RRF |
| Why is fixed-size chunking usually wrong? | Chunking |
| Why separate retrieval evaluation from generation evaluation? | Evaluating RAG |
| What limits how many users a GPU can serve? | KV cache and batching |
| An engineer has missed three commitments. Run the conversation. | Missed commitments |
| Explain zero trust and its sequencing to a director. | Zero trust |
| Push, pull or hybrid for a news feed? | Design a news feed |
| Why not just store a balance? | Design a payments ledger |
How to practise
Read only the question. Ninety seconds, out loud, standing, as if to a person. Then read the written answer and note the one thing you missed. Move on; do not re-answer immediately, because you will recite rather than recall.
Record yourself once a week and watch at 1.5x. Note filler words, rambling, and the moment you lost the interviewer. Fix one thing per session.
When you genuinely do not know: state the boundary, reason from adjacent knowledge, name how you would find out. Never bluff, because at this level interviewers probe two layers past your claimed knowledge specifically to find it, and one caught bluff outweighs several strong answers. And never stop at "I don't know" without the reasoning attempt, which is a wasted at-bat.
Source: §25, §34, §38.8, §39.7, §49.