Performance　optimization　especially　for　fault　tolerance　optimization　has　been　a　significant　aspect　for　in-memory　computing　computing.　When　node　failure　occurs,　in-memory　computing　may　lose　the　data,　which　increases　the　execution　time　without　checkpoint.　In　the　traditional　Spark　strategy,　the　programmer　chooses　the　checkpoint　with　the　uncertainty　and　risk.　Therefore,　we　aims　at　the　checkpoint　strategy　of　in　memory　computing　framework　Spark　in　this　paper.　After　the　theoretical　analysis,　the　checkpoint　selection　algorithm　which　taking　into　account　the　length　of　the　RDD　lineage,　the　computational　cost,　the　operation　complexity　and　the　size　in　setting　the　checkpoint　is　presented.　The　greater　the　weight　of　RDD,　the　higher　priority　it　has.　The　RDD　with　higher　cost　will　be　set　as　the　checkpoint　first,　which　can　reduce　the　recomputation　cost　of　the　task.　When　failure　occurs,　the　recovery　algorithm　is　executed,　and　the　efficiency　of　the　task　recovery　can　be　effectively　improved.　And　the　experimental　results　show　that　the　strategy　optimizes　the　fault　tolerance　mechanism　for　Spark　and　improves　the　efficiency　of　the　job　recovery.　©　2018,　Springer-Verlag　GmbH　Germany,　part　of　Springer　Nature.