Release, recovery and deferral of outbox batches still cost a write per row

未关闭
#3,410 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。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
52/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
csharp

调研方向

Start by reading OutboxDrainer.ReleaseAsync and the claim logic it uses for leased rows; then trace how each lane reads PluginState and builds the deferred-id exclusion list. Check the existing outbox tests before changing behavior. Done means each proposal item is fixed or explicitly declined in the PR, with tests covering every fix.

由索引模型根据 Issue 内容生成。

描述

area:events performance

Problem

These are left over from the review of PR #3404 (outbox lanes per owner). None of them loses an event; they cost writes, time or log noise.

  1. Release on shutdown, row by row and without a token. A stopping host releases each unclaimed row of a batch with its own conditional UPDATE (OutboxDrainer.ReleaseAsync), and the update takes no cancellation token. A large batch can therefore hold shutdown for BatchSize writes.
  2. Recovery after a failed release is row by row. A release that fails leaves the row leased with its claim attempt counted. Once the lease lapses, the row is claimed alone because it still names an owner. A whole failed batch is then recovered one claim per row.
  3. Deferrals cost a write each while plugin state is down. When PluginState cannot be read, every row of a batch is released and deferred with one write each. Each later drain then repeats the read of the plugin state for every lane.
  4. Deferrals are bounded by time only. A deferred row is skipped for MissingHandlerDelay, and the set of deferred ids is not capped. Their ids also go into every claim as a NOT IN parameter list. A long thread-pool starvation can make that list long.

Proposal

  1. Release the rest of a batch with one UPDATE ... WHERE id IN (...), bounded by a short timeout.
  2. After a lapsed lease, claim alone only the first row that names a dead owner, not each of them.
  3. Read plugin state once per lane pass, and on failure release the batch in one write.
  4. Cap the deferred set, for example at BatchSize, and leave the oldest out of the NOT IN list.

Acceptance

  • Each item fixed or explicitly declined in its PR, with tests for those fixed
主要语言
C#
星标
567
派生
107
平均合并
1 小时 19 分钟
30 天内合并 PR
353

环境准备

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。

fluentcms/FluentCMS 的其他 Issue

查看 fluentcms/FluentCMS 的全部 Issue

相似的 Issue

更多 C# Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。