The invention discloses a task scheduling method of a Spark cluster system by online partition optimization, and belongs to the technical field of online cluster resource scheduling. The method comprises the following steps that: carrying out statistics on an upstream Stage output data size, a total CPU (Central Processing Unit) nuclear number which participates in operation execution, a total memory amount and a memory proportion used for pulling data; according to a relationship between the output data size and memory capacity used for pulling data, calculating a task execution round number,and setting an optimal partition amount; monitoring the average CPU use ratio and the memory use ratio of a calculation node, and evaluating the resource use level of each calculation node; carryingout descending sort on the resource use levels of all nodes, and preferentially scheduling the task to be executed by the node with the highest use level; and repeating the above steps until all tasksfinish being scheduled. By use of the task scheduling method, an optimal partition amount can be automatically configured, the resource use ratio of the cluster is improved, and the execution speed of a Spark operation is quickened.