The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 84, in _split_generators
raise ValueError(
...<2 lines>...
)
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
IMDB database, query workloads, and DuckDB execution traces
The data of dbsched, pipeline-level query scheduling for DuckDB: the IMDB database of the Join Order Benchmark (JOB), SQL workloads over it, and the traces of running those workloads on a DuckDB fork that records how each query's pipelines executed.
The repository mirrors data/dataset/imdb/ of a dbsched checkout. python -m scripts.data.hub pull dataset downloads it there and unpacks the traces.
Layout
db/job/raw/schematext.sql, schema.json the IMDB schema
db/job/active/active.db the scale-1 database (DuckDB file, 2.3 GB)
db/job32/samples/<table>.parquet 1,000 sampled rows of each scale-32 table
queries/raw/<workload>.sql one query per line
queries/job32/<workload>/plans/trace/ traces on the scale-32 database
chunk_<n>.tar.gz up to 500 queries each
queries/job32/mscn/profiles/<n>_threads/ solo profiles of 5,000 mscn queries
profiles.parquet, trace/chunk_<n>.tar.gz
simulator/job32/mscn-5000/ concurrent runs of the same 5,000 queries
pool.json, <run>/trace/chunk_<n>.tar.gz
Database
db/job/active/active.db is the IMDB snapshot that JOB uses (21 tables, the May 2013 dump), loaded into DuckDB 2.0. A dbsched.meta table records its scale factor, 1.
Larger databases are not stored here: they are built from this one, inside DuckDB, in a few minutes. The traces were made on scale 32 (76 GB):
python -m scripts.db.scale_db --scale 32 --target data/dataset/imdb/db/job32/active/active.db
db/job32/samples/ holds 1,000 randomly sampled rows of each table of the scale-32 database, in the row order that matters: PQueryFormer's sample bitmaps record which of a table's sampled rows pass a query's filters, so the published weights expect these samples. scale_db loads them into the database as <table>_sample.
Workloads
| Workload | Queries | Plans traced on scale 32 |
|---|---|---|
job |
113 | 113 |
job-light |
70 | 70 |
job-ext |
24 | 24 |
job-d |
8,692 | 8,692 |
ceb |
13,563 | 13,561 |
job-light-scaled |
10,414 | 10,414 |
job-ext-scaled |
11,063 | 11,063 |
mscn |
105,318 | 105,317 |
brad |
10,000 | not traced |
ceb and job-d are de-duplicated. Of the 149,257 queries run, 149,254 have a plan. The 5 queries that failed (4 in ceb, 1 in mscn) are listed under errors in their chunk's run.json.
Traces
Each chunk_<n>.tar.gz unpacks to a chunk_<n>/ directory:
| File | Contents |
|---|---|
run.json |
the run's settings, the ids of its queries, and its errors |
queries.parquet |
one row per query |
pipelines.parquet |
the pipelines of each query |
tasks.parquet, intervals.parquet, ticks.parquet, attachments.parquet |
what the executor's threads did, and when |
plans/query_<id>.json |
DuckDB's profiler plan of the query, with operator timings |
A query's id is its line number in queries/raw/<workload>.sql, from 1. The trace tables are described in docs/ of the dbsched repository.
The runs: one query at a time (1 connection), 16 threads, a 100 ms tick interval, DuckDB's memory_limit at 38.1 GiB, on a VM with 16 vCPUs (Intel Sapphire Rapids) and 58 GB of memory. CPU times are specific to that hardware. DuckDB was the fork's jacob/simulator-data branch, based on v2.0.0 (2.0.0.dev2610201602).
Solo profiles and simulator runs
simulator/job32/mscn-5000/pool.json lists a pool of 5,000 mscn queries, sampled at random (seed 6616). They are traced in two ways, with the same chunk layout as above:
- Solo profiles,
queries/job32/mscn/profiles/{2,8}_threads/: each query run alone (1 connection) on 2 or 8 threads, with a 100 ms tick interval, in 10 chunks of 500.profiles.parquetholds the per-pipeline profiles derived from the trace (21,746 pipelines). - Simulator runs,
simulator/job32/mscn-5000/<policy>-c<connections>-t<threads>-tick<interval>-s<seed>/: the pool run concurrently over several connections, on 16 threads, while a scheduling policy (decisions every 20 ms) shares the threads among the running queries' pipelines. Ticks are every 1 ms. 20 chunks of 250 queries each.
| Run | Policy | Connections | Seed |
|---|---|---|---|
random-c4-t16-tick1000us-s1 |
random | 4 | 1 |
random-c8-t16-tick1000us-s0 |
random | 8 | 0 |
random-c16-t16-tick1000us-s2 |
random | 16 | 2 |
equal-c8-t16-tick1000us-s0 |
equal | 8 | 0 |
Query 135 crashed in every simulator run. It is listed in each run's trace/crashed.txt, so each run has 4,999 queries. The hardware, DuckDB build, and memory limit are those of the plan traces above.
Licence
The IMDB data is subject to IMDb's terms for non-commercial use. The query workloads come from the papers that introduced them (JOB, JOB-light, JOB-extended, MSCN, CEB, and others); check their terms before redistributing.
- Downloads last month
- 38