The Tale of Errors in Microservices (Artifact part 1)

This dataset provides comprehensive microservice traces (more than 1.5 million) collected from Uber microservice architecture, as described in our paper The Tale of Errors in Microservices, presented at SIGMETRICS 2025. This dataset enables researchers to study microservice behaviors, optimize performance, and investigate latency reduction techniques.

The data has been sanitized to protect proprietary information while retaining critical performance characteristics for academic research.

Artifact Structure and Decompression Instructions:

Due to Zenodo's file size constraints and upload issues, the large trace1-sanitized.tar.zst and trace2-sanitized.tar.zst files have been split into multiple pieces. The artifact is available in two parts (10.5281/zenodo.13947828 and 10.5281/zenodo.13952897). To access the sanitized microservice traces, download all the split parts from both artifacts. After downloading, reassemble the files using the following commands and then decompress the .zst files individually. Each .zst file will require 300-500GB of disk space to decompress.

Reassembling the split files:

For trace1-sanitized.tar.zst and trace2-sanitized.tar.zst, use the following commands to reassemble them:

cat trace1_* > trace1-sanitized.tar.zst

cat trace2_* > trace2-sanitized.tar.zst

Once reassembled, you can decompress the files:

zstd -d trace1-sanitized.tar.zst

zstd -d trace2-sanitized.tar.zst

Contents of the Traces:

trace1-sanitized.tar.zst and trace2-sanitized.tar.zst (in 10.5281/zenodo.13952897) contain over 1.5 million traces that correspond to the data described in Sections 3 and 4 of the original paper. Note: The traces in this dataset were collected on different days than those used in the paper, so analysis results may vary slightly from what is reported in the publication.

driver-sanitized.tar.zst contains the sanitized version of the original trace and corresponds to the App-Launch Use Case discussed in Section 6.3 and Figure 17 of the original paper.

Note

Due to privacy and security concerns, most unrelated fields and tags are removed. However, error-related tags are retained.

All trace is sanitized consistently. The same service or endpoint will have identical mapping across three directories. (i.e., service 1 represents the same service in all traces). However, the mapping is inconsistent with https://zenodo.org/records/13956078, so please do not mix the traces between the two artifacts.

To preserve privacy, the start time of each trace has been randomly shifted. As a result, the start and end times in the traces do not reflect the actual collection times, and users should not attempt to infer when the traces were gathered.

Within each trace, the relative durations and timestamps of all spans remained consistent, as the shift was applied uniformly across the entire trace.

If you use the traces in your research, please cite our paper

The Tale of Errors in MicroservicesI-Ting Angelina Lee (Washington University in St. Louis); Zhizhou Zhang, Abhishek Parwal (Uber Technologies Inc.); Milind Chabbi (Uber Technologies)SIGMETRICS 2025 https://doi.org/10.1145/3700436

If you have more questions, you can reach out to Chris(Zhizhou) Zhang.