learning-it by example

#323 Forward deployed engineer interviews 5.9 min

Estimate a Deployment: P50 and P90 With Monte Carlo

The plan said 34 days. One simulated project in a hundred finished that fast.

1. Why does the video simulate a deployment plan instead of adding up the most likely time of every task?
2. A made-up 12-task plan adds up to 34 days if every task takes its likely time. What did 10,000 simulated projects show?
3. What does the video mean by merge bias?
4. The video says the real spread of finish dates is likely wider than its simulation. Why?
5. How does the drivers function decide which tasks move the finish date most?

Practice only: nothing is saved and no score counts towards anything. Each answer comes with the reason.

  1. Shorten the likely time for accounts

    Copy the code folder first. In plan.py change the days of the 'accounts' task from (3, 7, 20) to (3, 4, 20). Predict the plan's total if every task takes its likely time, and say whether the chance of finishing within that many days will go up or down compared with the 1% for 34 days in the video. Run python run.py to check. Then change (3, 4, 20) back to (3, 7, 20), because the next exercise starts from the original.

    A hint

    Accounts sit on the longest chain of tasks, so every likely day you save comes off the total. I saw 31 days.

  2. Cap the worst case for accounts

    In plan.py change the days of 'accounts' from (3, 7, 20) to (3, 7, 10). Only the worst case changes. Predict whether the 90th percentile will fall and which task will become the top driver of the date. Run python run.py. Then change (3, 7, 10) back to (3, 7, 20), because the next exercise needs the original plan.

    A hint

    A smaller worst case lowers the spread and the accounts task stops dominating. I saw P90 of 49 days and the data feed as the top driver.

  3. Use fewer simulated projects

    In simulate.py change runs=10000 to runs=1000 in the simulate function. The printed text will still say 10,000, which is fine. Predict whether the percentiles will stay close to 46 and 54 days and whether the chance of finishing in 34 days will stay at 1%. Run python run.py.

    A hint

    A smaller sample is noisier, especially for rare events. I saw a chance of 0% and a different third driver.

These tasks are for you to try on your own machine. Nothing is checked or saved.

What this does not show

A made-up 12-task plan; the ranges and the triangular shape are assumptions; tasks are treated as independent, so the real spread is likely wider.

Up next in Forward deployed engineer interviews

More in Forward deployed engineer interviews

All Forward deployed engineer interviews videos

Related videos