为了解决传统以MapReduce为代表的批量计算在实时处理方面的瓶颈，流处理方法成为大数据处理的研究热点。然而海量数据流处理过程中的大量状态同步与性能干扰严重影响数据处理，离群数据增大故障误判概率。本项目从流计算开销模型、负载均衡、离群数据发现、故障恢复策略研究面向海量数据流处理容错，1）研究开销敏感弹性负载均衡，建立基于面积的非线性开销敏感模型与基于标准熵的均衡模型，减少高时延抖动，为流处理网络拓扑增加反馈，解决现有流处理网络无法运行迭代递归优化算法的问题；2）研究基于低水位滑动时间窗口的离群数据发现，在低水位滑动时间窗口内有效区分数据延迟到达与数据处理故障，减少故障恢复误判次数；3）研究基于批量状态依赖的并行上游备份故障恢复策略，减少故障恢复时层叠的重复计算、故障恢复状态重构时间和数据重放数量。该项目研究及其成果提高流处理的可用与可靠性，对推动流处理应用具有重要意义。

To address the bottleneck issues of MapReduce-based batch computing, stream-based processing becomes the research hotspot. However, large-scale state synchronization and performance interference will affect data processing seriously, and too many stragglers will increase the error probability. This study researches the fault tolerance of stream processing from the aspects of cost model, load balancing, straggler discovery, and fault recovery. This project will first research the cost-aware load balancing method, establish the area-based non-linear cost models to reduce the shake of high latency, and add the feedback loop to support iterative optimization algorithm in the stream-based topology. Second, this project will research the sliding time window method of low watermark to address straggler discovery and distinguish delay received from processing fault. Finally, this project will research the upstream mutual backup fault recovery based on batch state dependency to reduce the duplicated overlapping computation, the time to reconstruct the state, and the amount of data replay. This research and its achievements can improve the availability and reliability, and has important implications for stream processing applications.