Retry "already landed" check fails for an overwrite of an empty table, then deletes the landed snapshot's manifests

未关闭
  • 哪里不对？
#4,022 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

  • #4023 来自 @dancsi —— 已关闭，未合并

评估

难度
4/5
预计耗时
3-5 天
新手友好度
35/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python
领域
databases

调研方向

Start with Transaction.commit_transaction, especially the landed-snapshot check, and trace how Table.overwrite registers the delete and append producers. Use the reproduction in the issue to investigate the lost-response retry and cleanup behavior. Done means regression tests show the landed snapshot is recognized and its manifests remain available; the reporter says they already have a fix with tests.

由索引模型根据 Issue 内容生成。

描述

Apache Iceberg version: 0.12.0 (also reproduced on main at ebbc0ba3e4)

Please describe the bug 🐞

After a failed attempt, Transaction.commit_transaction checks whether the attempt actually landed:

if all(
    self._table.metadata.snapshot_by_id(producer._snapshot_id) is not None
    for producer in self._snapshot_producers
):

A transaction can contain a producer that never creates a snapshot. Table.overwrite on an empty table (or any delete that matches nothing, followed by an append) registers a delete producer, but only the append produces an AddSnapshotUpdate. The delete producer's _snapshot_id never appears in the table, so the check is always false, even when the append landed.

pyiceberg then rebuilds and retries. On the rebuilt attempt, the delete producer's concurrency validation sees the transaction's own landed append as a concurrent conflicting write, and raises ValidationException. That goes to _clean_all_uncommitted(), which deletes the manifest list and manifests of the snapshot that landed.

Reproduction
import tempfile

import pyarrow as pa
from pyiceberg.catalog.sql import SqlCatalog
from pyiceberg.exceptions import CommitFailedException
from pyiceberg.schema import Schema
from pyiceberg.types import LongType, NestedField

warehouse = tempfile.mkdtemp()
catalog = SqlCatalog("default", uri=f"sqlite:///{warehouse}/catalog.db", warehouse=f"file://{warehouse}")
catalog.create_namespace("default")
table = catalog.create_table(
    "default.t",
    Schema(NestedField(1, "a", LongType(), required=False)),
    properties={"commit.retry.min-wait-ms": "0"},
)

# The catalog applies the first commit, but the client sees a failure.
commit_table = catalog.commit_table
calls = 0
def commit_table_lose_first_response(*args):
    global calls
    calls += 1
    response = commit_table(*args)
    if calls == 1:
        raise CommitFailedException("simulated lost response")
    return response
catalog.commit_table = commit_table_lose_first_response

try:
    table.overwrite(pa.table({"a": pa.array([1, 2, 3], pa.int64())}))
except Exception as e:
    print(f"overwrite raised: {type(e).__name__}: {e}")

table = catalog.load_table("default.t")
snapshot = table.current_snapshot()
print("manifest list exists:", table.io.new_input(snapshot.manifest_list).exists())
table.scan().to_arrow()

Output:

Commit failed due to a concurrent update, retrying (1/4) in 0 ms
overwrite raised: ValidationException: Added data files were found matching the filter for snapshots set()!
manifest list exists: False
FileNotFoundError: [Errno 2] Failed to open local file '.../metadata/snap-...avro'

The same overwrite with a response that isn't lost succeeds, and a plain append with a lost first response is correctly detected as landed.

Suggested fix

Base the check on the snapshots the attempt actually tried to add, i.e. the AddSnapshotUpdates in self._updates, rather than on every producer. Equivalently, skip producers that didn't produce a snapshot. Ideally, a ValidationException raised during a rebuild should also never clean up a snapshot that is present in the refreshed metadata.

Related: #4021, which covers the missing check on the last attempt.

Willingness to contribute
  • I would be willing to contribute a fix for this bug with guidance from the Iceberg community

I have a fix with regression tests for both this and #4021: it keys the landed check on the AddSnapshotUpdate snapshot ids that were sent, and runs it after every CommitFailedException, including on the last attempt. I'm happy to open a PR.

主要语言
Python
星标
1.2k
派生
654
平均合并
2 天 11 小时
30 天内合并 PR
122

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 有 Pull Request 模板
  • 没有贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

apache/iceberg-python 的其他 Issue

查看 apache/iceberg-python 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。