Day 4 · Thu 2026-10-08 · ChatGPT CORE
ChatGPT CORE serving: streaming, reconnect, scheduling, quotas, and context cost
Week 11 trains two designs with one operating spine: admission, isolation, durable work, queue/backpressure, streaming/reconnect, derived state, evals, and recovery.
3contiguous 1-D DP slots
25-35mattempt + follow-up window
Staffinvariant, failure, recovery, metrics
Exact Daily Schedule
No drift: the same six blocks run every day.
| Time | Block | Contract |
|---|---|---|
| 08:50-09:00 | Target | Write one Staff-level invariant and one failure window before reading. |
| 09:00-10:05 | Project design | 25-35m ChatGPT serving follow-up, then streaming/scheduler/quota/context deep dives. |
| 10:05-10:45 | DDIA / source grounding | Map the design to exact DDIA anchors and only direct applied-AI sources. |
| 10:45-10:55 | Spoken close | Record the architecture in English, then say the Chinese invariant aloud. |
| 14:30-16:15 | 1-D DP contiguous block | Three NeetCode slots below; each slot is 25-35 minutes plus follow-up. |
| 20:30-21:15 | Recall / artifact | Run English Senior+/Staff recall and patch the scorecard evidence. |
Canonical Hello Interview Anchors
Direct day-page anchors only. Source titles were verified live before build.
| Canonical anchor | Disposition | Design decision |
|---|---|---|
| Potential Deep Dives | CORE deep dive | Deep dives must connect user-visible UX to serving constraints. |
| How do we stream tokens back fast and keep the stream smooth? | CORE deep dive | SSE cursor, replay buffer, heartbeats, and reconnect semantics. |
| How do we route and schedule generation requests across GPU workers? | CORE deep dive | Scheduler accounts for model, prompt length, tier, queue age, and GPU capacity. |
| How do we keep heavy users from monopolizing GPUs while giving paid tiers a better experience? | CORE deep dive | Quota, token buckets, queue lanes, and paid-tier fairness. |
| As conversations get longer, how do we control inference cost without making the assistant feel forgetful? | CORE deep dive | Context selection, summarization, retrieval, and prompt cache economics. |
| Cancelling a run and reclaiming the GPU | CORE deep dive | Cancellation is a run-state transition plus worker cleanup. |
Exact DDIA Mapping
Only Ch5, Ch9, and Ch13 sections are in scope.
| Chapter | Exact section | Interview use |
|---|---|---|
| Ch5 | 持久化执行 | Long model calls persist inputs, state, retries, and cancellation. |
| Ch5 | 消息代理 | Token events and status changes need replayable fanout for reconnect. |
| Ch9 | 超时和无界延迟 | GPU queues need explicit overload modes and tier fairness. |
Applied AI Source Gate
First-party/original sources appear only when they change a serving, eval, security, or operating decision.
| First-party/original source | Concrete decision it changes |
|---|---|
| OpenAI - Streaming API responses | Use event streams, cursors, and retryable UI behavior rather than treating the response as one blocking blob. |
| OpenAI - Background mode | Long-running generation or tool calls become durable background runs with polling or reconnect. |
| OpenAI - Prompt caching | Context-pack ordering and stable prefixes become serving and cost decisions. |
| OpenAI - Rate limits | Admission control must account for requests, tokens, tiers, and ingestion limits. |
21-Slot 1-D DP Runway
Today is slot 10-12 of 21, contiguous across the week.
| Block | NeetCode problem | Mode | Invariant | 25-35m follow-up | Complexity |
|---|---|---|---|---|---|
| 14:30-15:05 | Word Break | Prefix feasibility | dp[i] is true if any previous valid split plus dictionary word reaches i. | Ask for trie or max-word-length optimization. | O(n^2) / O(n) |
| 15:05-15:40 | Longest Increasing Subsequence | Patience tails | tails[k] is the smallest possible tail for length k+1. | Ask for O(n^2) DP proof and O(n log n) upgrade. | O(n log n) / O(n) |
| 15:40-16:15 | Partition Equal Subset Sum | 0/1 subset | Reverse iterate sums so each number is used once. | Ask why forward iteration corrupts 0/1 semantics. | O(n * target) / O(target) |
Artifacts And Recall
Use these in the evening block and scorecard.
ChatGPT Chinese deep lectureChatGPT Chinese lecture audioEnglish Senior+/Staff recall scriptEnglish recall PDF
Required spoken close: explain one invariant, one failure window, one recovery path, and one metric in English. Then restate the same design decision in Chinese without adding new components.
Detailed lecture notes, audio, recall scripts, PDFs, Staff Q&A, and mock packs are archived locally and are intentionally not published on this site.