An idle partition's active segment never seals, so its data never expires

オープン
  • 何か問題がありますか？
#4,650 コメント 3 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
5/5
見積もり時間
1週間以上
初心者へのやさしさ
25/100
issue の種類
機能追加
明瞭さ
おおむね明確
活発さ
活発
技術スタック
rust
領域
backend, distributed-systems

調査の方向性

Read Segment::is_expired in core/partitions/src/segment.rs, remove_sealed_segments_up_to in core/partitions/src/iggy_partition.rs, and the tiered storage RFC section linked in the issue. The issue leaves the solution open and describes why independent replica timers break segment placement, so first work out how a replicated RollSegment operation fits the existing control path. Done means idle active segments can seal without new batches and then participate in retention, compression, and offloading.

索引モデルが issue の本文から書いたものです。

説明

bug server
Description

A partition that stops receiving messages keeps its active segment open forever, and retention only ever acts on sealed segments. Segment::is_expired returns false for anything unsealed (there is a test for it unsealed_segment_is_never_expired), and remove_sealed_segments_up_to skips the active segment. So an idle partition holds that segment's data indefinitely, no matter what message_expiry or max_topic_size say. Segments it sealed before going idle still expire normally; it is the final partly filled one that never becomes eligible.

This came out of the discussion in #4402, which proposes a time ceiling on a segment's span but evaluates it on append and scopes idle partitions out. That split is the reason for a separate issue: a partition that still receives messages will eventually evaluate a trigger, whether that trigger is the current size check or a time check. A partition receiving nothing never evaluates one at all.

Where it shows up:

  • a decommissioned topic leaves every partition sitting on a partly filled segment that never seals and never expires
  • the same happens inside a live topic when traffic skews, since a partition that stops receiving keys is idle whether or not the topic is
  • an unsealed segment also cannot be compressed or offloaded, so this blocks tiering for those partitions
Affected area / component

Iggy server

Proposed solution

Left open.

Whatever closes this has to fire without an arriving batch, so it cannot be evaluated in the append path. The candidate raised in the tiered storage RFC is a replicated time roll: the primary appends a RollSegment control operation and every replica seals at that operation.

Alternatives considered
  1. Each replica rolling on its own timer. This does not work: segment boundaries have to be a function of the batches, and SegmentState::reserve runs on every replica, so a local now() would place the same batch differently on different replicas and fail the placement check in validate_segment_prefix.
Contribution
  • I'm willing to submit a pull request to implement this feature
Good first issue
  • I think this could be a good first issue for a new contributor
主要言語
Rust
スター
5.2k
フォーク
502
平均マージ
3日 3時間
マージ済み PR（30日）
163

環境構築

Codespaces で開く

このプロジェクトの開発コンテナを、あなたの GitHub アカウントでブラウザ上に起動します。

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。

apache/iggy のほかの issue

apache/iggy の issue をすべて見る

似ている issue

Rust 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。