Bekerja dengan engine streaming dan GPU

Menskalakan dan Mengoptimalkan Pipeline Data dengan Polars

Liam Brannigan

Data Scientist & Polars Contributor

Engine eksekusi query

Diagram alir membandingkan query dari API, melalui query optimizer ke in-memory engine.

Engine eksekusi query

Diagram alir membandingkan query dari API, melalui query optimizer ke in-memory engine.

Engine eksekusi query

Diagram alir membandingkan query dari API, melalui query optimizer ke in-memory dan streaming engine.

Engine eksekusi query

Diagram alir membandingkan query dari API, melalui query optimizer ke in-memory, streaming, dan GPU engine.

Ringkasan lazy request

requests = pl.scan_parquet("311_Service_Requests.parquet")

query = (
    requests
    .filter(pl.col("STATUS") == "Completed")
    .group_by("DEPARTMENT")
    .len()
    .sort("len", descending=True)
)

Menggunakan engine default

query.collect()

shape: (5, 2)
| DEPARTMENT        | len     |
| ---               | ---     |
| str               | u32     |
|-------------------|---------|
| 311 City Services | 4859161 |
| Sanitation        | 3406631 |
| Aviation          | 2337842 |
| Transportation    | 1468240 |
| Water             | 416535  |

Menargetkan engine streaming

query.collect(
    engine="streaming"
)

shape: (5, 2)
| DEPARTMENT        | len     |
| ---               | ---     |
| str               | u32     |
|-------------------|---------|
| 311 City Services | 4859161 |
| Sanitation        | 3406631 |
| Aviation          | 2337842 |
| Transportation    | 1468240 |
| Water             | 416535  |

Menargetkan engine streaming

Gambar tabel besar yang di-streaming.

Langkah streaming yang tidak didukung otomatis kembali ke engine default

Menargetkan engine streaming

import polars as pl
pl.Config.set_engine_affinity("streaming")

Menargetkan engine GPU

query.collect(
    engine="gpu"
)

Memerlukan perangkat keras GPU NVIDIA

Memilih engine

Mulai dengan engine default; beralih hanya jika memori terbatas

Streaming untuk query besar yang melebihi memori

GPU untuk kecepatan, hanya jika ada perangkat NVIDIA

Memilih engine

Bekerja per batch

completed_requests = (
    requests
    .filter(pl.col("STATUS") == "Completed")
    .select("SR_NUMBER", "TYPE", "DEPARTMENT", "CREATED_DATE")
)

Memproses tiap batch

for batch in completed_requests.collect_batches(

):

Memproses tiap batch

for batch in completed_requests.collect_batches(
    chunk_size=50_000,
):

Memproses tiap batch

for batch in completed_requests.collect_batches(
    chunk_size=50_000,
):
    print(batch.shape)
    batch_json = batch.write_json()

(50000, 4)
(50000, 4)
...
(50000, 4)
(12409, 4)

Ayo berlatih!

Menskalakan dan Mengoptimalkan Pipeline Data dengan Polars