docs: record access tiering introduction and retirement decisions

Signed-off-by: Feng Ruohang <rh@vonng.com>
This commit is contained in:
Feng Ruohang
2026-09-15 11:26:45 +08:00
parent 13bf126ebd
commit 4093fa0d78
3 changed files with 128 additions and 4 deletions
+8 -2
View File
@@ -24,9 +24,13 @@ and [complete commit range](https://github.com/pgsty/silo/compare/RELEASE.2026-0
- Reconcile ordinary single-object version DELETE across all pools, including - Reconcile ordinary single-object version DELETE across all pools, including
null versions, delete markers and unqualified directory-marker DELETE. This null versions, delete markers and unqualified directory-marker DELETE. This
prevents movement leftovers from surviving a successful response. Unreadable applies the deletion to every resolved pool copy under existing quorum
rules. Pending outbound delete replication retains versions until the
existing replication worker completes their purge; a successful response
does not imply immediate physical removal from every drive. Unreadable
pools now consistently return 503 instead of depending on pool traversal pools now consistently return 503 instead of depending on pool traversal
order; this extends the existing failure surface. Retry after recovery. order; insufficient read quorum returns `SlowDownRead`. This extends the
existing failure surface. Retry after recovery.
Cleanup failures also return an error. Batch deletion already fans out across Cleanup failures also return an error. Batch deletion already fans out across
pools; replication and scanner cleanup keep their existing contracts. See pools; replication and scanner cleanup keep their existing contracts. See
[scope and limitations](docs/bucket/lifecycle/access-tiering-removal.md#version-deletion-scope). [scope and limitations](docs/bucket/lifecycle/access-tiering-removal.md#version-deletion-scope).
@@ -35,6 +39,8 @@ and [complete commit range](https://github.com/pgsty/silo/compare/RELEASE.2026-0
its tracker, mover, scanner hooks, configuration, XML actions and metrics. its tracker, mover, scanner hooks, configuration, XML actions and metrics.
Accept and ignore retired configuration/XML and preserve ordinary statistics Accept and ignore retired configuration/XML and preserve ordinary statistics
when reading v9 caches. See [migration notes](docs/bucket/lifecycle/access-tiering-removal.md). when reading v9 caches. See [migration notes](docs/bucket/lifecycle/access-tiering-removal.md).
The [decision record](docs/investigations/access-tiering-revert.md) preserves
the feature's introduction, subsequent fixes, rollback scope and review history.
- Preserve the independent multi-pool write, metadata, healing and conditional - Preserve the independent multi-pool write, metadata, healing and conditional
deletion fixes from PR #178, including shared remote-tier reference protection. deletion fixes from PR #178, including shared remote-tier reference protection.
- Enforce `If-Match` on DELETE, preserve retention and independently ordered - Enforce `If-Match` on DELETE, preserve retention and independently ordered
@@ -2,6 +2,8 @@
PR #60 introduced an opt-in scheduler that moved objects between local server pools according to GET frequency. It and its feature-specific fixes have been removed. This does not remove ordinary lifecycle expiration, transitions to remote tiers, rebalance, decommission, or the general multi-pool correctness fixes from PR #178. PR #60 introduced an opt-in scheduler that moved objects between local server pools according to GET frequency. It and its feature-specific fixes have been removed. This does not remove ordinary lifecycle expiration, transitions to remote tiers, rebalance, decommission, or the general multi-pool correctness fixes from PR #178.
The [introduction and rollback record](../../investigations/access-tiering-revert.md) documents the commit history, scope decision, review corrections and unresolved validation findings.
The published Server 20260903 predates this feature. These instructions concern main/snapshot deployments that included #60; upgrading from the published version does not require access-tier configuration cleanup. The published Server 20260903 predates this feature. These instructions concern main/snapshot deployments that included #60; upgrading from the published version does not require access-tier configuration cleanup.
## Before upgrading a build with access tiering ## Before upgrading a build with access tiering
@@ -27,8 +29,8 @@ Interrupted rebalance/decommission can leave the same version in more than one p
## Version deletion scope ## Version deletion scope
Ordinary single-object `DELETE ?versionId=...` reconciles the addressed UUID, null version or delete marker across pools. Unqualified DELETE of a directory marker (a key ending in `/`) also addresses its null version and uses this path. A successful response means the addressed copies were removed; other versions remain. Ordinary single-object `DELETE ?versionId=...` reconciles the addressed UUID, null version or delete marker across pools. Unqualified DELETE of a directory marker (a key ending in `/`) also addresses its null version and uses this path. A successful request applies the deletion to every resolved pool copy under the existing per-pool quorum rules; other version IDs remain. If outbound delete replication is pending, copies retain `VersionPurgePending` until the existing replication worker completes the purge. Success does not guarantee immediate physical removal from every drive.
If a pool is unreadable, these requests can return 503 even when another pool has a readable copy. This extends an existing failure surface: previously the result could depend on whether the unreadable pool preceded the successful pool in traversal order; it now fails consistently. Retry after recovery. Cleanup failures also return an error. Ordinary unqualified DELETE retains its existing semantics. Batch `DeleteObjects` already fans out across pools. If a pool is unreadable, these requests can return 503 even when another pool has a readable copy. Insufficient read quorum returns `503 SlowDownRead`; other failures retain their corresponding error codes. This extends an existing failure surface: previously the result could depend on whether the unreadable pool preceded the successful pool in traversal order; it now fails consistently. Retry after recovery. Cleanup failures also return an error. Ordinary unqualified DELETE retains its existing semantics. Batch `DeleteObjects` already fans out across pools.
Incoming replicated deletes, lifecycle expiration, free-version cleanup and movement-internal calls retain their existing contracts. In particular, an incoming replicated version delete can leave movement duplicates in other pools; this change does not solve that separate case. Expiration scanners process their own pools and may remove duplicate expired copies in later cycles; free-version cleanup remains local to a pool. Do not treat the ordinary DELETE repair as a guarantee for every source of deletion. Incoming replicated deletes, lifecycle expiration, free-version cleanup and movement-internal calls retain their existing contracts. In particular, an incoming replicated version delete can leave movement duplicates in other pools; this change does not solve that separate case. Expiration scanners process their own pools and may remove duplicate expired copies in later cycles; free-version cleanup remains local to a pool. Do not treat the ordinary DELETE repair as a guarantee for every source of deletion.
@@ -0,0 +1,116 @@
# 访问频率分层:引入、修复与回退记录
记录日期:2026-09-15。本记录说明 PR #60 的设计、合入后的取舍,以及此次回退为什么同时保留并补齐通用多池正确性修复。操作步骤见[退役迁移说明](../bucket/lifecycle/access-tiering-removal.md)。合并、正式发布和生产部署是不同状态;本记录随回退变更交付,不代表已经发布。
初稿审查时(2026-09-15 03:20 UTC),最终候选尚未移植到远端基线、尚未冻结 PR head,合并前全量检查与三次有效 Linux 运行尚未执行,本变更尚未合并。后续执行状态以承载本记录的 PR 及其绑定提交的验收记录为准;下文的历史实验不替代这些检查。
## 1. 引入的目标和实际范围
[@mrjavadseydi](https://github.com/mrjavadseydi) 在 [PR #60](https://github.com/pgsty/silo/pull/60) 提出了基于 GET 频率的本地池间分层。成功 GET 更新有界滚动计数,后台调度器把热对象提升到配置中的首个池,把已经迁移且变冷的对象降到末个池。功能默认关闭,至少需要两个池;它与普通生命周期过期、远端对象存储 transition、rebalance 和 decommission 是不同机制。
实现不只是一个后台任务:它增加了 GET 计数入口、leader 调度、跨池版本栈复制、来源清理及失败恢复、热池配额、生命周期 XML 扩展、配置项、指标和扫描统计。访问热度统计使 data-usage cache 从 v8 升到 v9。对象本体的存储格式没有因此改变。
搬移要保持完整版本历史、删除标记、null version、时间戳、ETag、校验和和加密元数据;同时需要处理并发写入、目的端已有版本、来源部分删除和远端 tier 引用。合入前的修补和测试针对并覆盖了这些边界,不应把回退解释成贡献无效。贡献者署名继续保留,独立的其他贡献也不回退。
## 2. 可追溯时间线
下表的 PR/Issue 时间使用 UTC;提交链接对应具体代码,不把“报告时间”当作“缺陷首次出现时间”。
| 时间 | 事件 | 本次处理 |
| --- | --- | --- |
| 2026-08-15 10:15 | #60 创建;原始实现 [`7a060cab1`](https://github.com/pgsty/silo/commit/7a060cab1edd5bbc17da7f703bbd1ab7415b6f7c) | 随特性撤销 |
| 2026-09-06 00:09 | [#133](https://github.com/pgsty/silo/issues/133) 报告多池副本写不能权威调和 Object Lock 状态 | 保留解决它的通用修复 |
| 2026-09-06 15:12 | [#144](https://github.com/pgsty/silo/issues/144) 报告条件 DELETE 原子性仅限单个纠删码集合 | 保留解决它的通用修复 |
| 2026-09-08 05:18 | [`9a6e1477f`](https://github.com/pgsty/silo/commit/9a6e1477f45067559def8423d431ee177795134f) 补访问分层兼容标识清单 | 删除功能专属标识,保留有依据的退役兼容 |
| 2026-09-08 07:08 | [`374de0fa3`](https://github.com/pgsty/silo/commit/374de0fa32aa1d6eda57dbba6ac522ebf793b6be) 修复搬移的版本保全、写隔离和删除范围 | 随专属搬移器撤销 |
| 2026-09-08 07:28 | [`5ac33e158`](https://github.com/pgsty/silo/commit/5ac33e1583e838ade3f56c7d60e80e7a854f9a88) 确定性覆盖搬移失败恢复 | 随已删除搬移器的专属测试撤销 |
| 2026-09-08 07:42 | #60 以 [`a3df317ae`](https://github.com/pgsty/silo/commit/a3df317ae0725eb650e4d3e21551154f69be6229) 合入 | 以该 merge 的第一父差异确定功能边界 |
| 2026-09-11 12:34 | [#178](https://github.com/pgsty/silo/pull/178) 合入通用多池写入、元数据与条件删除调和 | 保留,解除测试对访问搬移器的依赖 |
| 2026-09-13 | [`2dd1e00da`](https://github.com/pgsty/silo/commit/2dd1e00da49faf995f2db29807fc6211b8376d7d) 的 CHANGELOG 同时汇总访问分层和通用多池修复 | 拆开表述,不整条删除独立修复历史 |
| 2026-09-15 | 维护者决定收缩访问分层;完成来源分析、三种候选反证、退役兼容、普通版本 DELETE 修补与外部评审 | 形成此次选择性回退 |
#178 中的 [`e59a3d938`](https://github.com/pgsty/silo/commit/e59a3d938ed25c1bcd51efbb4ad6955073d195f7) 提供池级串行化、字段调和和条件删除;[`ccb676e60`](https://github.com/pgsty/silo/commit/ccb676e60cb7441ee65ff7c35f3b7828979101fd) 保护仍被其他副本使用的远端 tier 引用;[`51d41345f`](https://github.com/pgsty/silo/commit/51d41345f7ac532f9c6fea2b7dba5f29da930b8f) 保存 Linux 重启及 OIDC 验收记录。这三项不属于仅为访问频率调度而存在的代码。
截至回退评估时,公开 Server `RELEASE.2026-09-03T13-18-01Z` 早于 #60 合入。退役迁移主要针对运行过后续 main、自行构建或快照版本的实例;不能据此声称正式 release 用户普遍启用过此特性。
## 3. 为什么回退,为什么不能整批撤销后续修复
维护者的取舍是:默认关闭的可选调度能力,对配置、生命周期、缓存、统计和核心多池写入路径带来了过大的维护面。此次移除的是该能力及专属实现,没有测量并宣称吞吐量提升、延迟降低或固定减少一次分布式锁往返。
“后来改过同一文件”不等于“由 #60 引发”。#133/#144 的报告早于 #60 合入;更关键的是,原有 rebalance、decommission 和复制写入也会使同一版本暂时存在于多个池。访问分层消失后,这些状态仍然合法存在。移除调和与锁纪律会重新允许旧副本遮蔽新元数据、条件删除选错版本、清理错误被吞掉等问题。
评估在隔离工作树中实际比较了三条路线:
| 候选 | 通用多池回归 | 普通指定版本 DELETE |
| --- | --- | --- |
| A:撤销 #60,保留 #178 | 原有 13 组通过 | 仍能成功返回后留下可读副本 |
| B:同时撤销 #60 和 #178 的存储修补 | 相同 13 组中 10 组失败 | 问题仍在 |
| C:A 加普通版本 DELETE 调和 | 13 组原有及当时新增的 8 组通过 | 同一复现通过 |
这些是 2026-09-15 的历史对照结果,不是最终 PR head 的发布验收。后续补上目录标记、真实 rebalance 中断等覆盖后,通用多池测试达到 23 组。测试通过不能替代来源分析,来源分析也不能替代最终候选的运行验证。
## 4. 最终保留与删除的边界
- 删除访问 tracker、调度/搬移器、GET 和 scanner 钩子、热池配额、专属配置帮助、生命周期动作、指标及专属测试。
- 保留普通过期、远端 transition、rebalance/decommission、复制写入,以及 #178 的 Object Lock、标签、条件删除、元数据调和与远端引用保护。
- 原有 data-usage 生成解码器恢复到 #60 前的实现;允许读取 v8/v9,利用字段编码跳过已退役热度字段,继续写 v8。普通字段保真由历史真实 v9 样本测试覆盖。
- 仅容忍准确的十个退役 ILM 键;读取生命周期时丢弃退役扩展。纯访问动作的规则需先清理才能再次编辑;混合规则保留普通动作。
- 已经搬移的对象留在当前池;没有全量搬回、自动删除所有重复版本、后台清理服务或对象元数据重写。
曾对本地审查提交 `d06f1c614` 做过声明来源复核:消失的 357 个声明均不在 #60 之前,删除的十个文件均由 #60 引入;#60 原先删换的 41 行旧文本按忽略空白比较有 40 行恢复,剩下一行保留 #178 在已持有池锁时调用 `getWritePoolIdx(..., true)` 的修正,避免对同一对象再次取锁。生成缓存解码器与功能前逐字节一致,原有 13 组通用测试没有删除。这是该审查版本的保全证据,不能把数字脱离 SHA 当作未来所有版本的保证。
## 5. 普通版本 DELETE 是独立补洞
功能移除不会自动消除历史重复副本。普通单对象指定版本 DELETE 因而复用既有调和路径:在池级对象锁内读取每个池的目标版本,计算一次条件及回调,先删除非权威副本,再处理权威副本;任何不可读池或清理错误都不能当作成功。
范围包括 UUID、null version、delete marker,以及原先就被解析为 null version 的未指定版本目录标记 DELETE。入站复制、搬移内部调用、生命周期过期和 free-version 清理保留各自语义;批量 `DeleteObjects` 原本就会向池并发扇出,不是此次遗漏。
删除标记需要向 retention/metadata 回调传入与 set 层相同的 `MethodNotAllowed` 或 `ObjectNotFound` 语义。直接复用拒绝 marker 的元数据更新入口会错误地拒绝合法版本删除。回调从所有副本合并独立更新的 Object Lock 和标签,不能随意只采用一个池的状态。
存在两项明确的成功/失败边界:
1. 读法定多数不足时返回 `503 SlowDownRead`,即使另一个池有可读副本。旧路径的结果会受池遍历顺序影响;新路径把失败语义统一。它是正确性与可用性的取舍,需要恢复后重试。
2. 出站删除复制尚未完成时,成功响应可以表示各副本进入 `VersionPurgePending`,由既有 worker 完成清理。原有每池 quorum 规则也继续适用;不能把成功响应等同于每一块盘立即物理删除。
## 6. 审查如何改变了方案
本地先形成三个线性审查提交:`8fdfdabd9` 移除特性,`d06f1c614` 补普通 DELETE,`6e3fdca97` 去除重复扫描。它们记录审查演进,最终 PR 在独立远端基线上重放,提交 ID 会改变;不应把线性演进误认为三份同时维护的实现。
Claude Code Opus 5 / max 的五轮实现评审要求补齐退役兼容、说明协调停机及环境一致性、验证真实池故障,并纠正运行证据措辞。随后 Claude 与 ZCode 的独立复核再次确认了回退边界和普通 DELETE 语义;最终两轮计划商榷收束了交付流程。
| 意见 | 裁定与处理 |
| --- | --- |
| 指定版本 DELETE 重复扫描所有池 | 接受。第一次扫描已在同一锁内得到目标副本,直接合并其结果。16 盘池在回调前的读取计数从 32 降为 16;这不等于总 I/O 或延迟减半。 |
| N 个副本产生 N 条 DELETE 审计 | 反驳。底层调用只追加上下文标签,HTTP 层向每个配置审计目标发送一次请求事件。成功完成调和时池标签最终指向 primary;不增加额外 NoAuditLog 修改。 |
| retention/metadata 顺序与 set 层不同 | 差异存在,但已有明确注释。保留 retention 优先,拒绝后不删除、不调度删除复制、不 Sweep;不宣称任意自定义回调都能交换。 |
| 把优化 amend 进旧提交并直接丢弃 main 脏修改 | 不 amend 已审历史。先保存完整文件、补丁、哈希及可达恢复引用,核对覆盖后受控恢复。八组新增通用测试承接,访问搬移专用测试由真实 rebalance 中断覆盖替代。 |
| 从当前本地分支直接开 PR | 调整。其祖先含另一个任务的 IAM/超时提交 `ebc9937d9`;从远端 `89637554d` 仅移植本次变更,不夹带或删除独立工作。 |
| 合并之后再跑全量与 Linux 验收 | 不接受。先完成文档和移植,冻结 PR head,再验收;不能把旧 SHA 的测试直接提升为新基线的通过记录。 |
| 不同基线 diff 必须逐字节一致 | 改为每提交 stable patch-id、路径与完整树等价性核验。blob hash 和行号随基线变化,不应成为错误的拒绝依据。 |
`objectPoolInfos` 的并行查询作为独立性能跟进,本次不增加并发实现。Contributor 署名、#132 配额指标、#77 桶元数据、federated COPY、IAM、超时及其他独立修复不因本次取舍被整体撤销。
## 7. 历史运行证据与未定案事项
以下是移植前 `d06f1c614` 的历史实验,不是最终 PR head 的验收替身。
| 运行 | 实际观察 | 不应推导的结论 |
| --- | --- | --- |
| `f590538f`,四节点双池 | 旧版真实 rebalance 中断留下 9 个重复 UUID;协调停机换新后对象四节点可读,旧配置与普通规则可编辑,真实缓存头 v9→v8;删除一个 addressed UUID 后四节点 HEAD/GET 404,另一个版本仍可读 | 没有验证全部 9 个不同 UUID 收敛;物理元数据仅每池抽查一盘;没有整份运行时统计守恒测量 |
| `fd93cd37`,三节点双池 | 停掉一个池所在节点,保留 namespace 锁的 2/3 法定多数;DELETE 返回 503 SlowDownRead,源全部四盘保留目标版本;恢复后 DELETE204,各节点 HEAD/GET404 | 不能推广为跨主机网络、持久盘及压力验收 |
| 被弃用的四节点停池拓扑 | 同时丢失 namespace 锁法定多数,发生客户端超时,记录脚本还遇到 NoneType 错误 | 不是“池读取返回503”的证明 |
| `83676ca2` | 升级后配置/生命周期编辑之后一次 HeadObject 返回503;artifact 没有记录 DELETE 自身响应 | “DELETE之后”仅来自脚本顺序,不能写成已证实 DELETE204 后异常,也不能归类为已修复、既有问题或暂态 |
**开放项:`83676ca2` 的 HEAD503 仍未定案。** 可直接比较的目标阶段历史运行是一失败、一成功;一次未复现不足以关闭问题。仅凭 `SlowDownWrite` 等错误名字也不能给其他失败确定容量或环境根因。
最终候选的合并前复核采用有界规则:要求三次有效升级后 DELETE/HEAD 运行,最多五次总尝试;每次记录 DELETE 码/耗时、失败 HEAD 的节点与版本、GET 错误码、两池全盘元数据、固定间隔重试时序和旧版同拓扑对照。旧版可能保留副本返回200,不要求它满足新增跨池删除契约。
有证据证明在目标操作前失败的 harness 尝试才能不计入有效运行,但仍计入总尝试。任何目标阶段的新503或数据不变量失败都不能通过补跑抹掉,必须暂停合并并定位。三次通过也只满足这项工程检查,不证明历史异常已消失;开放项在合并和正式发布评估时仍须可见。
## 8. 交付与恢复纪律
最终候选使用独立分支;main 的五个旧修改先保存完整内容、二进制补丁、SHA-256 和具名 Git 恢复引用,再核对原 HEAD/哈希及测试覆盖,只恢复这五个文件。禁止用整树 reset 或 clean 代替受控归一;若用户已有新增编辑,应保留并重新核对。
全量 cmd/internal、相关 race、构建、vet、lint、生成文件和兼容检查,以及上述 Linux 验收,绑定最终候选的实际 SHA。若纳入新的远端提交,重新记录基线与 head,复核变更并重跑受影响验收。文档与原始测试记录各自保留其对应版本,不篡改旧失败、不用新通过覆盖旧记录。
逐节点滚动升级未通过既有二进制校验检查,采用[协调停机方案](../bucket/lifecycle/access-tiering-removal.md#before-upgrading-a-build-with-access-tiering)。实验使用单个 Docker Linux VM 和 tmpfs;正式 tag、包、镜像、跨主机及生产部署仍是独立交付。未验证全部重复 UUID 或运行时统计守恒应如实披露,不能反向引入自动搬回/清理需求或无关重构。