Construction of highly reliable distributed computing model in multi-source heterogeneous data environment

In order to meet the requirements of high reliability and low latency in multi-source heterogeneous data scenarios, this paper proposes a highly reliable distributed computing model based on "data-computing-fault tolerance". Firstly, the heterogeneous data adaptation layer is designed, and the lossless fusion of cross-domain data is realized by dynamic format conversion and ontology alignment mechanism, which reduces the preprocessing delay to 90 ms and improves the throughput to 5.8 GB/s; Secondly, an adaptive fault-tolerant framework based on reinforcement learning (RL) is constructed. Through LSTM fault prediction and Q-learning task migration, the task success rate is still 96% under the 15% node failure rate, and the average recovery time is only 2.1s; Finally, the game theory hybrid resource scheduler is introduced to balance the heterogeneous CPU/GPU/FPGA load with the node utility function, and the Pareto optimization of throughput and delay is realized. The 32-node cluster experiment shows that the end-to-end delay of the model is 4.5× lower than that of Spark, and the delay fluctuation under fault impact is less than 8%, which can support millisecond-level key tasks such as remote surgery and industrial control.