Cloud ML platforms let you trade cost for reliability by choosing how many training instances are cheap-but-revocable spot capacity versus full-price on-demand capacity. This simulator runs a synchronized data-parallel training cluster where spot instances can be reclaimed at random, rolling the whole job back to its last checkpoint. Tune the spot share, interruption rate, checkpoint interval and restart time to see how the mixed fleet's total cost and completion time compare live against an all-on-demand baseline.