rohitg00--ai-engineering-from-scratch
wehub-resource-sync
91 行
3.4 KiB
JSON
91 行
3.4 KiB
JSON
{
|
|
"lesson": "34-repo-memory-and-state",
|
|
"title": "Repo Memory and Durable State",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "What is the durability test that decides whether a piece of information belongs in repo memory?",
|
|
"options": [
|
|
"Whether it is JSON",
|
|
"Whether it would be useful three months from now in a CI rerun; if yes, repo; if no, telemetry",
|
|
"Whether the user marked it as important",
|
|
"Whether it fits in 4 kB"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Repo memory is for durable, three-months-from-now-useful state; transient data is telemetry."
|
|
},
|
|
{
|
|
"stage": "pre",
|
|
"question": "What schema field carries the agent's contract version?",
|
|
"options": [
|
|
"schema_version",
|
|
"build_hash",
|
|
"model_id",
|
|
"session_uuid"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "schema_version is the integer contract; the manager refuses to load from an unknown version."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "How does atomic write work?",
|
|
"options": [
|
|
"tempfile.mkstemp in the same directory, write, fsync, os.replace (atomic rename) over the target",
|
|
"Truncate-then-write to the target",
|
|
"Append-only with a CRC",
|
|
"Encrypt and overwrite"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "Atomic rename on POSIX and Windows is what prevents partial-write corruption."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Why are idempotency keys required for non-idempotent tool calls?",
|
|
"options": [
|
|
"For billing",
|
|
"If the agent crashes after a tool call but before checkpointing the result, retry safely; log call ID before execution and skip the call on retry",
|
|
"To deduplicate logs",
|
|
"To shorten output"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "pending_calls.jsonl carries the call IDs; recovery checks and skips already-executed work."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Where should large artifacts (CSVs, long transcripts, generated files) live relative to state?",
|
|
"options": [
|
|
"Inline in agent_state.json",
|
|
"As separate files (or object storage) with only the path kept in state, so checkpoints stay small and fast",
|
|
"In environment variables",
|
|
"Concatenated into one giant log"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Separate artifacts grow independently of state; checkpoints stay cheap to read and write."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What does event sourcing for audit + snapshots for resume buy you?",
|
|
"options": [
|
|
"Lower disk usage",
|
|
"Replay agent decisions verbatim by reading the snapshot then replaying events after it; same shape as Postgres WAL",
|
|
"Faster inference",
|
|
"Native voice support"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Append to state.events.jsonl on every mutation; periodically snapshot to state.json; replay events after the snapshot timestamp."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What does the lesson say happens when schema_version mismatches?",
|
|
"options": [
|
|
"The manager silently upgrades",
|
|
"The manager refuses to load until a migration script in tools/migrate_state.py runs",
|
|
"The state is deleted",
|
|
"The agent retries"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Schema migrations or refuse-to-load; never silent upgrade."
|
|
}
|
|
]
|
|
}
|