Project Planning & Execution Flashcards

A data science team is debating whether to build a custom feature store or use an existing open-source solution. Which factor should MOST heavily influence this build-vs-buy decision?

Answer: The team's scale of feature reuse across projects, latency requirements, and available engineering bandwidth relative to the open-source solution's capabilities

Build-vs-buy decisions for ML infrastructure should be driven by fit-gap analysis of requirements versus existing capabilities, weighted by the engineering cost of building and maintaining a custom solution.

What is the role of a 'pre-mortem' in ML project planning?

Answer: A proactive exercise where the team imagines the project has already failed and works backward to identify the most likely causes

A pre-mortem primes the team to surface hidden risks before the project begins by using prospective hindsight to identify failure modes that might be dismissed in forward-looking planning.

When managing a long-running ML project, 'model staleness' refers to:

Answer: Performance degradation caused by shifts in the real-world data distribution relative to the distribution the model was trained on

Model staleness occurs when the world changes after training, causing the input distribution or label relationship to drift and eroding model performance over time.

Which project execution strategy is MOST effective for managing uncertainty in novel ML research tasks?

Answer: Time-boxing exploration phases with clear decision gates that determine whether to continue, pivot, or stop based on early evidence

Time-boxed exploration with pre-defined decision gates balances innovation with accountability, preventing indefinite research without delivery while leaving room for genuine discovery.

A team is planning to retrain an ML model on a monthly schedule. Which condition should trigger an UNSCHEDULED retraining between cycles?

Answer: A statistically significant data drift or model performance degradation detected by monitoring alerts

Monitoring-triggered retraining responds to actual performance signals rather than arbitrary schedules, ensuring the model is updated when needed rather than on a fixed cadence that may be too slow or too frequent.

In ML project resource planning, 'GPU memory' is a hard constraint because:

Answer: Batch size, model size, and optimizer states must collectively fit within VRAM or training will fail with out-of-memory errors

GPU VRAM must hold model parameters, activations, gradients, and optimizer states simultaneously; exceeding it causes OOM errors that halt training entirely, not just slow it down.

Which artifact should be produced at the END of every ML sprint to ensure project continuity if a key team member leaves?

Answer: A sprint summary documenting experiment results, key decisions made and their rationale, and updated project assumptions

Sprint summaries capturing experiment outcomes, decision rationale, and revised assumptions create institutional memory that allows new team members to onboard quickly and prevents repeated mistakes.