What are NOT adopted by the Bosen system to balance communication overhead with convergence of data$-$parallel machine learning algorithms?

$($A$)$ Upon failures, it simply restarts from scratch

$($B$)$ Execution guarantees similar to bulk synchronous parallelization cannot be ensured

$($C$)$ It selects a middle$-$ground consistency model called bounded staleness consistency model

$($D$)$ It uses clock messages to ensure bounded staleness.