Kuiz berkaitan panduan · Sederhana Tahap
Kuiz Strategi Rollback dan Fallback Model
Bezakan beralih kepada model terdahulu daripada kembali kepada tingkah laku berganti, dan kenal pasti keperluan keserasian dan pencetus.
soalan 1 daripada 8
Bagaimanakah rollback berbeza daripada fallback?
Pratonton setiap soalan8 soalan
Jawapan dan penjelasan muncul semasa anda mengambil kuiz.
- Bagaimanakah rollback berbeza daripada fallback?
- Rollback restores a prior release; fallback uses an alternate behavior
- Rollback changes model outputs while fallback restores Git
- They are identical strategies in all systems
- Fallback always retrains the model
- Isyarat yang manakah boleh mencetuskan rollback operasi?
- A model's training accuracy only
- A sustained increase in serving errors or timeouts
- A color change in a dashboard
- An unreviewed comment in source code
- Mengapa mengekalkan artifak terdahulu yang tidak boleh diubah?
- It contains every request log
- It guarantees no future data drift
- It identifies the known-good bytes to restore
- It replaces health monitoring
- Apakah yang boleh menghalang model lama daripada berfungsi selepas pemulangan semula?
- A signed artifact digest
- A lower training loss in the candidate
- An alert dashboard
- An incompatible feature or data schema change
- Mengapa menguji tingkah laku sandaran secara bebas?
- Fallbacks cannot affect user decisions
- A fallback can have its own failure modes and safety tradeoffs
- A simpler model is always safer
- Testing would disable rollback
- Why use a measurement window and hysteresis for automated rollback?
- To hide real errors
- To delay every response indefinitely
- To reduce reactions to transient noise and repeated switching
- To guarantee the model passes evaluation
- What can rollback not reverse automatically?
- A model version reference
- Future traffic routing
- The artifact selected by the deployment controller
- Irreversible side effects already caused by prior predictions
- What should a release runbook include?
- Triggers, owner, recovery steps and fallback limitations
- Only the model's parameter count
- A guarantee of zero user impact
- The training set's class names