metadata

language: en
tags:
  - log-analysis
  - hdfs
  - anomaly-detection
license: mit
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: event_encoded
      dtype: string
    - name: tokenized_block
      sequence: int64
    - name: block_id
      dtype: string
    - name: label
      dtype: string
    - name: __index_level_0__
      dtype: int64
  splits:
    - name: train
      num_bytes: 1159074302
      num_examples: 460048
    - name: validation
      num_bytes: 145089712
      num_examples: 57506
    - name: test
      num_bytes: 144844752
      num_examples: 57507
  download_size: 173888975
  dataset_size: 1449008766

HDFS Logs Train/Val/Test Splits

This dataset contains preprocessed HDFS log sequences split into train, validation, and test sets for anomaly detection tasks.

Dataset Description

The dataset is derived from the HDFS log dataset, which contains system logs from a Hadoop Distributed File System (HDFS). Each sequence represents a block of log messages, labeled as either normal or anomalous. The dataset has been preprocessed using the Drain algorithm to extract structured fields and identify event types.

Data Fields

block_id: Unique identifier for each HDFS block, used to group log messages into blocks
event_encoded: The preprocessed log sequence with event IDs and parameters
tokenized_block: The tokenized log sequence, used for training
label: Classification label ('Normal' or 'Anomaly')

Data Splits

Training set: 460,049 sequences (80%)
Validation set: 57,506 sequences (10%)
Test set: 57,506 sequences (10%)

The splits are stratified by the Label field to maintain class distribution across splits.

Source Data

Original data source: https://zenodo.org/records/8196385/files/HDFS_v1.zip?download=1

Preprocessing

We preprocess the logs using the Drain algorithm to extract structured fields and identify event types. We then encode the logs using a pretrained tokenizer and add special tokens to separate event types. This dataset should be immediately usable for training and testing models for log-based anomaly detection.

Intended Uses

This dataset is designed for:

Training log anomaly detection models
Evaluating log sequence prediction models
Benchmarking different approaches to log-based anomaly detection

see honicky/pythia-14m-hdfs-logs for an example model.

Citation

If you use this dataset, please cite the original HDFS paper:

@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 I},
  booktitle={Proceedings of the ACM SIGOPS 22nd symposium on Operating systems principles},
  pages={117--132},
  year={2009}
}