apache / apache/iceberg-python

FsspecFileIO: `_adls` mutates shared properties, so a second storage account gets the first account's filesystem

オープン 初心者向け
#3,885 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る
主要言語
Python
スター
1.1k
フォーク
581
平均マージ
1日 17時間
マージ済み PR(30日)
77

説明

### Apache Iceberg version

main (development)

### Please describe the bug 🐞

`_adls` writes the account name it infers back into the properties dict it is handed, and the call site passes the FileIO's own `self.properties`. So the first ADLS location a `FsspecFileIO` touches pins `adls.account-name` for the life of that FileIO, and every later location is served a filesystem built for the first account.

In `pyiceberg/io/fsspec.py`:

```python
def _adls(properties: Properties, hostname: str | None = None) -> AbstractFileSystem:
...
# Fallback: extract account_name from URI hostname
if hostname and ADLS_ACCOUNT_NAME not in properties:
properties[ADLS_ACCOUNT_NAME] = hostname.split(".")[0]
```

and the caller at `fsspec.py:515`:

```python
if scheme in _ADLS_SCHEMES:
return _adls(self.properties, hostname)
```

The `lru_cache` on `(scheme, hostname)` just above it is not the problem. It does correctly build a second filesystem for a second hostname. But by then `ADLS_ACCOUNT_NAME` is already present in the shared properties, so the inference is skipped and the second filesystem gets the first account.

### Steps to reproduce

Two locations in two different storage accounts, starting from empty properties:

```python
from unittest.mock import patch
from pyiceberg.io.fsspec import FsspecFileIO

loc_a = "abfss://data@accountone.dfs.core.windows.net/wh/t/a.parquet"
loc_b = "abfss://data@accounttwo.dfs.core.windows.net/wh/t/b.parquet"

captured = []
class FakeFS:
def __init__(self, **kw): captured.append(kw.get("account_name"))

io = FsspecFileIO(properties={})
with patch("adlfs.AzureBlobFileSystem", FakeFS):
io.new_input(loc_a)
print(dict(io.properties))
io.new_input(loc_b)
print(dict(io.properties))

print(captured)
```

Output:

```
{'adls.account-name': 'accountone'}
{'adls.account-name': 'accountone'}
['accountone', 'accountone']
```

Expected `['accountone', 'accounttwo']`.

### What I expected

A FileIO given no `adls.account-name` should infer the account per location, not once. The inference is already per hostname at the cache layer, only the write into the shared dict breaks it.

The same mutation happens a few lines above in the SAS token loop, which sets `ADLS_ACCOUNT_NAME` from the token key, so that path has the same effect.

### Impact

A catalog whose tables span two storage accounts silently reads from the wrong account. Depending on whether a same named container exists there, this either fails with a confusing not found or resolves to the wrong data. It also means a FileIO's properties change as a side effect of reading, which is surprising for anything that inspects or reuses them.

### Suggested fix

Do not write into `properties`. Resolve the account into a local value inside `_adls` and pass that to `AzureBlobFileSystem`, leaving the caller's dict untouched. Happy to send a PR.

### Willingness to contribute

- [x] I can contribute a fix for this bug independently
- [ ] I would be willing to contribute a fix for this bug with guidance from the Iceberg community
- [ ] I cannot contribute a fix for this bug at this time

コントリビューションガイド

このリポジトリのコントリビューションガイドは索引されていません

調査の方向性

pyiceberg/io/fsspec.py の _adls と、515 行目付近にあるその呼び出し元から始め、次にモックされた AzureBlobFileSystem を使って 2 か所での再現を実行します。各 hostname がそれぞれ独自のアカウントを生成することと、2 回の読み取り後も io.properties が変更されていないことを確認します。

索引モデルが issue の本文から書いたものです。

評価

技術スタック
python
領域
backend, cloud
issue の種類
バグ
難易度
2/5
見積もり時間
1〜3時間
活発さ
活発
明瞭さ
明確に書かれている
初心者へのやさしさ
78/100

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。