shawhin/HDFS_v1_blocks

HDFS v1 Block-Level Dataset

Resources:

• Video Explainer
• GitHub Repo
• Dataset

Dataset Summary

This dataset is a block-level transformation of the HDFS_v1 log dataset. While the original dataset contains individual log lines (~11M rows), this version aggregates all log entries belonging to the same block into a single text sequence, making it suitable for LLM-based anomaly classification.

Each row represents a unique HDFS block with its complete log history concatenated in chronological order.

Supported Tasks

• anomaly-detection: Binary text classification to predict whether a block experienced an anomaly based on its log sequence.

Dataset Structure

Data Splits

Split	Examples	Normal	Anomaly
train	460,048	446,578	13,470
dev	57,506	55,822	1,684
test	57,507	55,823	1,684

Data Fields

Field	Type	Description
block_id	string	Unique HDFS block identifier (e.g., blk_-1608999687919862906)
text	string	Concatenated log entries for the block, newline-separated
label	int	Binary anomaly label (1 = anomalous, 0 = normal)

Text Format

Each log line within text follows the format:

<LEVEL> <COMPONENT>: <CONTENT>

Example:

INFO dfs.DataNode$DataXceiver: Receiving block blk_-1608999687919862906 src: /10.251.73.220:42557 dest: /10.251.73.220:50010
INFO dfs.DataNode$DataXceiver: Receiving block blk_-1608999687919862906 src: /10.251.73.220:55213 dest: /10.251.73.220:50010
INFO dfs.FSNamesystem: BLOCK* NameSystem.allocateBlock: /mnt/hadoop/mapred/system/job_200811092030_0001/job.jar. blk_-1608999687919862906
INFO dfs.DataNode$PacketResponder: PacketResponder 1 for block blk_-1608999687919862906 terminating
INFO dfs.DataNode$PacketResponder: Received block blk_-1608999687919862906 of size 67108864 from /10.251.73.220

Source Data

• Original Dataset: logfit-project/HDFS_v1
• Original Source: LogPAI/loghub

Dataset Creation

1. Loaded the original line-level HDFS_v1 dataset
2. Formatted each log line as <LEVEL> <COMPONENT>: <CONTENT>
3. Grouped by block_id and concatenated log entries (ordered by line number)
4. Aggregated anomaly labels (max per block)
5. Created stratified 80/10/10 train/dev/test splits preserving class distribution

Citation

@inproceedings{xu2009detecting,
  title={Detecting Large-Scale System Problems by Mining Console Logs},
  author={Xu, Wei and Huang, Ling and Fox, Armando and Patterson, David and Jordan, Michael},
  booktitle={SOSP 2009}
}

@inproceedings{zhu2023loghub,
  title={Loghub: A Large Collection of System Log Datasets for AI-driven Log Analytics},
  author={Zhu, Jieming and He, Shilin and He, Pinjia and Liu, Jinyang and Lyu, Michael R.},
  booktitle={ISSRE 2023}
}

License

See original dataset license.