Hacktoberfest 2026：メンテナが10月に向けて印を付けた、オープンで初心者向けの issue。 Hacktoberfest の issue を見る

Reusing intermediate results causes memory issues

オープン
  • 何か問題がありますか？
#854 コメント 6 件 リアクション 0 件 担当者 1 名 GitHub で見る

@phofl がすでに取り組んでいます。

2024年2月14日 から。

評価

この issue はまだ評価されていません。

説明

Problem

Whenever we reuse intermediate results and there is a pipeline breaker (such as shuffles, joins, reductions, or groupby operations), it forces us to materialize the entire intermediate result (thus, it is breaking the pipelining that we could utilize for reuse for example between multiple element-wise operations).

This result materialization puts a hard limit on our ability to scale as I have observed in multiple TPC-H benchmark queries.

To illustrate this, run these two snippets on a cluster of your choice

With full intermediate result materialization

from dask_expr.datasets import timeseries

from distributed import Client

if __name__ == "__main__":
    with Client() as client:
        print(client.dashboard_link)
        df = timeseries(start="2000-01-01", end="2020-12-31", freq="100ms", dtypes={"x": float})
        # To compute mean, we have to fully materialize df, and we won't free its data
        # until we have reused the chunks to compute a partial of the sum. 
        mean = df["x"].mean()
        df[df["x"] > mean].sum().compute()

Without full intermediate result materialization

from dask_expr.datasets import timeseries

from distributed import Client

if __name__ == "__main__":
    with Client() as client:
        print(client.dashboard_link)
        df = timeseries(start="2000-01-01", end="2020-12-31", freq="100ms", dtypes={"x": float})
        # Compute the mean beforehand so that we don't have to keep all of `df` in memory
        mean = df["x"].mean().compute()
        df[df["x"] > mean].sum().compute()

Possible solution

The easiest approach would be to never reuse any intermediate results. This has a few downsides:

  • Non-deterministic functions will lead to unexpected results
  • We waste a lot of computational resources on recomputations

...but it will allow us to scale.

We can certainly get smarter about intermediate result materialization, but this will require some effort depending on how smart we want to be. (There's a body of (ongoing) research and implementations in the database world we could draw from.)

主要言語
Python
スター
89
フォーク
26
30日以内にマージされた PR はありません

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

dask/dask-expr のほかの issue

dask/dask-expr の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。