QuitoBench — TsFile

Converted to Apache TsFile format from the original Hugging Face dataset hq-bench/quitobench. The dataset description below follows the original card. License: CC-BY-4.0 (same as source). Values are unchanged; this is a format conversion only.

QuitoBench is a regime-balanced evaluation benchmark curated from Quito, a billion-scale, single-provenance time series dataset of application-traffic workloads from Alipay's production platform.

🌐 Project Page: hq-bench.github.io/quito 📄 Paper: arXiv:2603.26017 💻 Code: github.com/alipay/quito 📦 Training Corpus: hq-bench/quito-corpus

Dataset Overview

	hour config	min config
Granularity	1 hour	10 minutes
# test series	517	773
Series length	15,356 steps	5,904 steps
Test-set length / series	552 steps	3,312 steps
Date range	2021-11-18 → 2023-08-19	2023-07-10 → 2023-08-19
# variates / series	5	5
Total rows	7,939,052	4,563,792

The 1,290 test series are stratified across all eight trend × seasonality × forecastability (TSF) regime cells (~160 series/cell), ensuring balanced evaluation.

Train/test split (upstream): Global temporal cutoff at 2023-07-28 00:00:00 UTC. The data published here is the test set.

Source Schema

Each row represents one timestamp of one series (long/tidy format).

Column	Type	Description
item_id	int64	Unique series identifier
date_time	datetime64[ns] (UTC)	UTC timestamp
ind_1 … ind_5	float64	Five anonymised traffic variates

TsFile Layout

.
├── README.md
├── hour/
│   ├── test_hour_1.tsfile … test_hour_8.tsfile   (table: quitobench_hour)
└── min/
    ├── test_min_1.tsfile  … test_min_5.tsfile    (table: quitobench_min)

Each config is one TsFile table, written by the converter as several shards (the tooling splits large outputs into multiple .tsfile parts). All shards of a config carry the same table schema; read them together to get the full table.

• Table: quitobench_hour / quitobench_min
• TAG (device): item_id — the source int64 id rendered as a STRING (TsFile tags must be STRING), e.g. "100011". One series = one device.
• Time: date_time (UTC) → INT64 epoch milliseconds.
• FIELD: ind_1, ind_2, ind_3, ind_4, ind_5 → DOUBLE (source float64).

No columns were dropped; both test splits have no nulls and no (item_id, date_time) duplicates, and every series has a uniform length.

Reading the TsFiles

import glob
from tsfile import TsFileReader

# Iterate over all shards of one config and concatenate
total = 0
for shard in sorted(glob.glob("hour/*.tsfile")):
    reader = TsFileReader(shard)
    name = next(iter(reader.get_all_table_schemas()))          # "quitobench_hour"
    schema = reader.get_all_table_schemas()[name]
    # Include the TAG column (item_id) explicitly to get it back in the result.
    cols = [c.get_column_name() for c in schema.get_columns()]  # item_id, ind_1..5
    with reader.query_table(name, cols, batch_size=200_000) as rs:
        while (batch := rs.read_arrow_batch()) is not None:
            df = batch.to_pandas()   # columns: time, item_id, ind_1..ind_5
            total += len(df)
print(total)  # 7,939,052 for hour

Note on querying tags. A TAG column (item_id) is a device-identifying dimension. To get it back in query results, include it explicitly in the column list. A query of only the tag (no FIELD column) returns 0 rows — this is expected TsFile table-model behavior, not a corrupt file.

License & Citation

Released under CC BY 4.0. Please preserve attribution to the original authors.

• Original dataset: hq-bench/quitobench

@article{xue2026quitobench,
  title   = {{QuitoBench}: A High-Quality Open Time Series Forecasting Benchmark},
  author  = {Xue, Siqiao and Zhu, Zhaoyang and Zhang, Wei and
             Cai, Rongyao and Wang, Rui and
             Mu, Yixiang and Zhou, Fan and Li, Jianguo and Di, Peng and Yu, Hang},
  journal = {arXiv preprint arXiv:2603.26017},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.26017}
}