Spark achieves fault tolerance through various mechanisms. One key approach is the resilient distributed dataset (RDD), which allows Spark to keep track of the lineage of data transformations. RDDs are immutable and can be reliably reconstructed in case of failure. Additionally, Spark employs data replication and task scheduling to mitigate the impact of failures. It also leverages speculative execution by launching multiple copies of a task and using the first one that completes, thereby reducing the impact of straggler tasks.