Sample Parquet File Download

Which Sample Parquet File Should You Use?

How to Download, Inspect, and Query a Parquet File

The examples below use Python with pyarrow and pandas (pip install pyarrow pandas) and DuckDB. The output shown is what you get with the files on this page.

Step 1: Download the file

Click Download file next to any sample above, or fetch it from the command line. Each file has a stable direct URL, so you can also download it in a CI job.

# Download a single file
curl -O https://samples-files.com/samples/code/parquet/sample1.parquet

# Download the edge-case files
curl -O https://samples-files.com/samples/code/parquet/sample4-data-types.parquet \
     -O https://samples-files.com/samples/code/parquet/sample6-int96-timestamps.parquet \
     -O https://samples-files.com/samples/code/parquet/sample7-empty.parquet \
     -O https://samples-files.com/samples/code/parquet/sample8-corrupted.parquet

Step 2: Inspect the schema and metadata

A Parquet file describes itself. You can read the schema, row groups, compression, and column statistics from the footer without loading any data.

import pyarrow.parquet as pq

# The schema is stored in the file footer, so this reads no data
print(pq.read_schema("sample4-data-types.parquet"))
# customer_id: int32 not null
# email: string not null
# is_active: bool
# signup_date: date32[day]
# last_login: timestamp[us, tz=UTC]
# balance: decimal128(10, 2)
# score: double
# country: dictionary<values=string, indices=int8, ordered=0>
# tier: string
# tags: list<element: string>
# address: struct<street: string, city: string, zip: string>
# preferences: map<string, string ('preferences')>

# Row groups, compression, and min/max statistics
md = pq.ParquetFile("sample5-large.parquet").metadata
print(md.num_rows, md.num_row_groups, md.row_group(0).column(0).compression)
# 300000 6 ZSTD
stats = md.row_group(0).column(1).statistics
print(stats.min, stats.max)
# 2025-01-01 00:01:01 2025-03-02 22:00:30

Step 3: Load it into a DataFrame

import pandas as pd  # needs pyarrow installed

df = pd.read_parquet("sample1.parquet")
print(df.shape)  # (30, 7)
print(df.head(3))
#    EmployeeID     Name  Age  Salary Department    HireDate                                                   Skills
# 0         101    Alice   43   77101         HR  2023-08-12  [Excel, Recruitment, Data Analysis, Accounting, Python]
# 1         102      Bob   54  114508      Sales  2022-03-02                                     [SEO, Data Analysis]
# 2         103  Charlie   31   54247  Marketing  2017-03-02            [Python, Machine Learning, Content Marketing]

# Empty file: you still get the columns, just no rows
print(pd.read_parquet("sample7-empty.parquet").shape)  # (0, 12)

Step 4: Read only what you need

The main advantage of Parquet is that you don't have to read the whole file. Select columns and push filters down to the reader:

import pandas as pd
import pyarrow.parquet as pq

# Read only 3 of 7 columns, and only rows from December 2025.
# Row groups whose min/max statistics can't match are skipped.
table = pq.read_table(
    "sample5-large.parquet",
    columns=["order_id", "order_ts", "unit_price"],
    filters=[("order_ts", ">=", pd.Timestamp("2025-12-01"))],
)
print(table.num_rows)  # 25498

Or query the file with SQL in DuckDB:

-- DuckDB queries Parquet files directly, no import step
SELECT category,
       count(*) AS orders,
       round(sum(quantity * unit_price), 2) AS revenue
FROM 'sample5-large.parquet'
GROUP BY category
ORDER BY revenue DESC
LIMIT 3;

┌─────────────┬────────┬─────────────┐
│  category   │ orders │   revenue   │
│   varchar   │ int64  │   double    │
├─────────────┼────────┼─────────────┤
│ Toys        │  37867 │ 52354366.37 │
│ Sports      │  37418 │ 51975307.79 │
│ Electronics │  37563 │ 51933555.02 │
└─────────────┴────────┴─────────────┘

Step 5: Test the edge cases

Check that legacy INT96 timestamps come back as real timestamps:

import pyarrow.parquet as pq

# The column is stored with the deprecated INT96 physical type
print(pq.ParquetFile("sample6-int96-timestamps.parquet").schema.column(2).physical_type)
# INT96

# pyarrow converts it back to a normal timestamp
print(pq.read_table("sample6-int96-timestamps.parquet").schema.field("event_time"))
# event_time: timestamp[ns]

And confirm that a corrupted file raises an error you can catch:

import pyarrow as pa
import pyarrow.parquet as pq

try:
    pq.read_table("sample8-corrupted.parquet")
except pa.ArrowInvalid as e:
    print(e)
# Could not open Parquet input source 'sample8-corrupted.parquet':
# Parquet magic bytes not found in footer. Either the file is corrupted
# or this is not a parquet file.

# DuckDB reports the same problem differently:
# Invalid Input Error: No magic bytes found at end of file 'sample8-corrupted.parquet'