Bekerja dengan dataset multifile

Menskalakan dan Mengoptimalkan Pipeline Data dengan Polars

Liam Brannigan

Data Scientist & Polars Contributor

File harian 311

311_requests_daily_csv/
  311_requests_2025-12-31.csv
  311_requests_2026-01-01.csv
  311_requests_2026-01-02.csv
  311_requests_2026-01-03.csv
  311_requests_2026-01-04.csv

Memindai banyak file CSV

pl.scan_csv(


)

Memindai banyak file CSV

pl.scan_csv(
    "311_requests_daily_csv/311_requests_2026*.csv",
    try_parse_dates=True,
)

Memindai banyak file CSV

pl.scan_csv(
    "311_requests_daily_csv/311_requests_2026*.csv",
    try_parse_dates=True,
).select("REPORTED_AT", "STATUS", "WARD").collect()

shape: (4, 3)
| REPORTED_AT         | STATUS      | WARD |
| ---                 | ---         | ---  |
| datetime[us]        | str         | i64  |
|---------------------|-------------|------|
| 2026-01-01 07:10:00 | In Progress | 46   |
| 2026-01-01 08:25:00 | Open        | 46   |
| 2026-01-04 07:45:00 | Open        | 12   |

Menulis partisi hive

pl.read_csv(
    "311_Service_Requests.csv",
    try_parse_dates=True,
).write_parquet(


)

Menulis partisi hive

pl.read_csv(
    "311_Service_Requests.csv",
    try_parse_dates=True,
).write_parquet(
    "311_requests_parquet/",

)

Menulis partisi hive

pl.read_csv(
    "311_Service_Requests.csv",
    try_parse_dates=True,
).write_parquet(
    "311_requests_parquet/",
    partition_by=["CREATED_DATE"],
)

Dataset terpartisi

311_requests_parquet/
  CREATED_DATE=2025-12-31/
    data.parquet

Dataset terpartisi

311_requests_parquet/
  CREATED_DATE=2025-12-31/
    data.parquet
  CREATED_DATE=2026-01-01/
    data.parquet

Memindai dataset berpartisi hive

requests = pl.scan_parquet(
    "311_requests_parquet/",
    hive_partitioning=True,
)

Nilai partisi → kolom biasa

Kueri sadar partisi

requests.filter(
    pl.col("CREATED_DATE") >= date(2026, 1, 1)
)

Kueri sadar partisi

requests.filter(
    pl.col("CREATED_DATE") >= date(2026, 1, 1)
).group_by("DEPARTMENT").len().collect()

shape: (2, 2)
| DEPARTMENT     | len |
| ---            | --- |
| str            | u32 |
|----------------|-----|
| Transportation | 208 |
| Sanitation     | 175 |

Kueri tidak sadar partisi!

requests.filter(
    pl.col("YEAR") >= 2026
)

Polars memfilter langsung pada kolom partisi 💡
Memfilter pada kolom turunan tidak memicu optimasi 💡

Masalah perubahan skema

# file harian 2026-01-01
REPORTED_AT,STATUS,WARD,BLOCK_CODE,REPORT_TEXT
2026-01-01 07:10:00,In Progress,46,0021,...

# file harian 2026-01-02
REPORTED_AT,STATUS,WARD,REPORT_TEXT
2026-01-02 09:05:00,Open,29,...

Menyisipkan kolom yang hilang

requests = pl.scan_csv(
    "311_requests_daily_csv/311_requests_2026*.csv",
    try_parse_dates=True,
    missing_columns="insert",
)

Memeriksa skema gabungan

requests.select(
    "REPORTED_AT", "STATUS", "WARD", "BLOCK_CODE"
).collect()

shape: (8, 4)
| REPORTED_AT         | STATUS      | WARD | BLOCK_CODE |
| ---                 | ---         | ---  | ---        |
| datetime[us]        | str         | i64  | i64        |
|---------------------|-------------|------|------------|
| 2026-01-01 07:10:00 | In Progress | 46   | 21         |
| 2026-01-01 08:25:00 | Open        | 46   | 21         |
| 2026-01-02 09:05:00 | Open        | 29   | null       |
| 2026-01-02 10:42:00 | Completed   | 29   | null       |

Masalah ketidakcocokan dtype

# file harian 2025-12-31
WARD
35

# file harian 2026-01-03
WARD
"44"

Dua pemindaian dengan dtype berbeda

ward_int = pl.scan_csv(
    "311_requests_daily_csv/311_requests_2025-12-31.csv",
    try_parse_dates=True,
)

ward_str = pl.scan_csv(
    "311_requests_daily_csv/311_requests_2026-01-03.csv",
    try_parse_dates=True,
    schema_overrides={"WARD": pl.String},
)

Menggabungkan dengan vertical_relaxed

combined = pl.concat(
    [ward_int, ward_str],
    how="vertical_relaxed",
)

Supertype: dtype yang menampung nilai dari kedua input
Di sini, String dapat menampung nilai dari Int64 dan String

Memeriksa concat yang relaks

combined.collect()

shape: (4, 5)
| REPORTED_AT         | STATUS    | WARD | BLOCK_CODE | REPORT_TEXT                      |
| ---                 | ---       | ---  | ---        | ---                              |
| datetime[us]        | str       | str  | i64        | str                              |
|---------------------|-----------|------|------------|----------------------------------|
| 2025-12-31 16:20:00 | Open      | 35   | 17         | Massive pothole getting worse    |
| 2025-12-31 18:05:00 | Open      | 35   | 17         | Cyclists forced into traffic     |
| 2026-01-03 07:35:00 | Open      | 44   | 12         | How is this still not repaired   |
| 2026-01-03 08:15:00 | Completed | 44   | 12         | Every bus hit shakes the vehicle |

Ayo berlatih!

Menskalakan dan Mengoptimalkan Pipeline Data dengan Polars