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

from_dataset re-reads every dimension coordinate on each call, and the workaround loses the cftime UDF and template registry

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

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
3/5
見積もり時間
1〜2日
初心者へのやさしさ
45/100
issue の種類
機能追加
明瞭さ
おおむね明確
活発さ
静か
技術スタック
python, sql
領域
api, data

調査の方向性

Start in XarrayContext.from_dataset, where coord_arrays is built once per call, and in read_xarray_table, which already takes coord_arrays. Also check _registered_datasets and _maybe_register_cftime_udf. Done means one Dataset can register several per-variable tables with each dimension coordinate read once, while metadata recovery and the cftime() UDF still work. The issue proposes two API shapes and the maintainers have not picked one, so confirm the direction before coding.

索引モデルが issue の本文から書いたものです。

説明

Problem

XarrayContext.from_dataset materialises the dimension coordinates once per call:

coord_arrays = {
    str(dim): input_table.coords[dim].values for dim in input_table.dims
}

Within a single call those arrays are shared across every dim group, which is what the comment above them describes. Across calls nothing is shared. Registering one table per data variable, which is what you want when each variable should be its own SQL table, re-reads every coordinate once per variable.

Measured with a counting DataArray.values property, 4 data variables on a (time, lat, lon) grid:

4 x from_dataset (one per data var): 12 coord reads
1 x from_dataset (whole dataset)   : 3 coord reads

Local NumPy makes this free. On a Zarr store each read is a network round trip, so registration cost scales with n_vars * n_dims instead of n_dims.

The obvious workaround is not equivalent

read_xarray_table already accepts coord_arrays, so the natural caller-side fix is to read the coordinates once and register the tables directly:

shared = {str(d): ds.coords[d].values for d in ds.dims}
ctx = XarrayContext()
for v in ds.data_vars:
    ctx.register_table(v, read_xarray_table(ds[[v]], chunks, coord_arrays=shared))

That does share the reads, but bypassing from_dataset silently drops two things it does.

  1. self._registered_datasets[name] = input_table is never populated, so XarrayDataFrame loses metadata recovery. Checked: _registered_datasets is {} afterwards, and to_dataset() then returns empty attrs and cannot infer dims.

  2. _maybe_register_cftime_udf never runs, so the cftime() UDF is missing on a 360_day dataset:

cftime() escape hatch : ERR ValueError Error during planning: Invalid function 'cftime'. Did you mean 'to_time'?
cftime() from_dataset : 4 rows

So today there is no public way to register several tables out of one Dataset without either paying n_vars * n_dims coordinate reads or giving up metadata recovery and cftime support.

Possible directions

Either shape would fix it:

  • Accept coord_arrays on from_dataset (the parameter already exists one layer down in read_xarray_table), so a caller registering several tables from one parent Dataset can hand in the arrays it already read.
  • Add a per-variable split to from_dataset, alongside the existing dim-group split, so a single call can register air, hum and so on as separate tables from one Dataset. This keeps the coordinate sharing, the _registered_datasets tracking and the cftime UDF registration together in one place.

Parent: #126

主要言語
Python
スター
142
フォーク
19
平均マージ
1日 9時間
マージ済み PR（30日）
9

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。

xqlsystems/xarray-sql のほかの issue

xqlsystems/xarray-sql 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。