Skip to content

docs(acp): ACP 4.5 + Alauda OS 标准交付规划(单集群 / 多集群)— 中文审阅稿 - #853

Closed
chinameok wants to merge 17 commits into
mainfrom
docs/acp-standard-delivery-dcs
Closed

chinameok wants to merge 17 commits into
mainfrom
docs/acp-standard-delivery-dcs

Conversation

@chinameok

@chinameok chinameok commented Oct 12, 2026 •

Copy link
Copy Markdown
Contributor

这是 ACP 4.5 + Alauda OS 标准交付规划的中文审阅稿,覆盖单集群和多集群两种交付方式:要准备的机器、节点池与插件的放置和存储、按顺序排列的操作步骤(每一步链接到产品文档)。

  • 本 PR 删除了此前基于 ACP 4.4 的 5 篇交付文章,改为这一篇新文章。
  • 现在是中文,方便直接在 PR 上评审。评审通过后会改为英文,再转为 Ready for review。
  • ACP 4.5 GA 前不合并。
  • 文中链接指向 ACP 4.5 文档站;个别小节线上内容还落后于 acp-docs release-4.5 分支(例如 Next Steps 的「Publish the Common Extensions」)。
  • 磁盘已写入正文。
  • review/acp-s2-resource-requirements.zh.md 只用于 review 各节点池的资源规格,合并前删除。

🤖 Generated with Claude Code

The production delivery counterpart to Full-Stack Automation with ACP:
the reference design (component selection, topologies, node pools and
taint scheme, single persistent /var/cpaas disk, per-cluster
observability), the end-to-end delivery flow including the DR domain-
switch sequencing, the acp-delivery-charts rendering workflow, plugin
placement by pool layout, machine configuration order, sizing by
derivation from the capacity pages, and evidence-based acceptance.

Encodes decisions verified in real ACP 4.3 deliveries. Tracking:
AIT-74918 under Epic AIT-74911.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The link checker strips .md/.mdx but not .markdown, so the relative
link resolved to a non-existent .markdown.html route and failed the
build. Reference the article by title instead; titles are searchable
in this knowledge base.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The delivery charts are withdrawn: hand-orchestrating the Cluster API
manifests is the product reality on ACP 4.3, and no chart tooling will
be maintained for the released lines. Section 3 now points at the
product pages that hold the complete manifests (global install Step 4,
the DR increments, Huawei DCS "Using YAML"), and keeps only what those
pages scatter or deliveries get wrong: naming as a hard requirement on
both sides, credentials out of applied manifests, one manifest set per
cluster, reading shared values from global via the DCS page's
placeholder table, keeping the provider literals, maxSurge 0 for
persistent-disk pools, the three DR increments identical on both sides,
and a pre-apply checklist. Wherever a product page already states a
rule, the article links instead of restating it.

The sizing and archive sentences that referred to chart values or
rendered output are reworded to match.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The product documentation this article links to adopted the acp/specs
platform glossary (alauda/acp-docs#1187, alauda/immutable-infra-docs
#221/#222): "traditional OS" is now "non-Alauda OS" and "bootstrap
host" is "bootstrap node". Follow the same wording so a reader moving
between the article and the linked pages sees one name per concept.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gate

R&D confirmed that an Alauda OS global installation does not need two
bootstrap nodes — the primary and standby installations can share one.
The delivery flow now provisions one bootstrap node for both, the "never
reuse one bootstrap node" line is removed, and the single-cluster
topology no longer lists "one bootstrap node" as if it distinguished it
from the multi-cluster case.

Stage 5 also listed "CPU meets x86-64-v2" as a pass criterion for the
bootstrap node. The installation guide states that baseline applies to
cluster nodes, not to the bootstrap node; the criterion now points at
the guide's bootstrap node requirements instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Corrects the previous commit. R&D has now confirmed (2026-10-15) that a
disaster-recovery delivery of an Alauda OS global needs two bootstrap
nodes, one to install the primary and one to install the standby; the
shared-bootstrap reading was a misunderstanding. Stage 5 provisions two
again, "Never reuse one bootstrap node for both sides" is back, and the
single-cluster notes list one bootstrap node as what sets them apart.

Kept from the previous commit: the bootstrap node has no x86-64-v2 pass
criterion (that baseline applies to cluster nodes), and the wrapped
"bootstrap host" now reads "bootstrap node".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nifests

Replace the ACP 4.3 "orchestrate and link" DCS draft with five ACP 4.4
articles that let a delivery engineer complete a standard delivery by
filling in parameter files and applying the given manifests:

- How to Deliver ACP with Immutable Infrastructure (common): reference
  design, parameter files and the render helpers (site.env + one file
  per cluster, envsubst with an explicit variable list, Bash fallback),
  delivery flow, platform plugins (Machine Configuration, MetalLB,
  Ingress NGINX, VictoriaMetrics, ClickHouse + Keeper + Kafka logging),
  machine configuration, sizing and acceptance commands.
- Huawei DCS, Huawei Cloud Stack and Bare Metal: global primary and
  standby (disaster recovery) plus workload clusters.
- VMware vSphere: workload clusters only.

Every manifest is assembled from the ACP 4.4 product documentation,
uses ${VAR} placeholders documented in a parameter table, keeps
credentials out of files and command lines, and is followed by the
check that proves the step worked. The articles say they must be
validated in a staging environment before a production delivery.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@chinameok chinameok changed the title How to Deliver ACP on Huawei DCS with Immutable Infrastructure docs(acp): deliver ACP 4.4 on Immutable Infrastructure — common article + DCS / HCS / Bare Metal / vSphere with complete manifests Oct 15, 2026
Three corrections to the ACP 4.4 Immutable Infrastructure articles.

Local storage is TopoLVM. The reference design replaces the open "choose a
StorageClass whose volumes survive node replacement" guidance with the
selected system: the console-driven install, the node-local constraint the
product states (application-layer replication only, no independent
persistence guarantee, ReadWriteOnce), and the StorageClass values this
delivery chooses. Two gaps are stated rather than papered over: TopoLVM
needs one unformatted disk per node, which only vSphere documents how to
provide, and the product's `lvm2` prerequisite assumes a mutable node.

The logging stack install is deferred. Its deployment scheme is being
revised, so the manifests and checks are removed and the section now
points at the Logging Service documentation while keeping the placement,
storage and sizing decisions this design already fixes. The log pool, its
nodes and its disks stay in the provider manifests.

Disaster recovery no longer explains steps with "`cpaas-system` is
excluded from synchronization". The rule set works on key prefixes with a
maintained deny-list; most objects cross, Secrets included. Each `global`
side keeps its own plugins, because `ModuleInfo` objects named `global-`
are excluded, and its own machine configuration, which is not
synchronized at all — so machine configuration is applied on both sides.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…elivery plan (Chinese review draft)

Remove the five ACP 4.4 Immutable Infrastructure delivery articles and add one
article that plans the standard delivery of ACP 4.5 on Alauda OS for a single
cluster and for multiple clusters: machines, node pools, plugin placement and
storage, and the ordered steps, each linking to the product documentation.

The article is in Chinese for review and will be translated into English before
the pull request leaves Draft.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@chinameok chinameok changed the title docs(acp): deliver ACP 4.4 on Immutable Infrastructure — common article + DCS / HCS / Bare Metal / vSphere with complete manifests docs(acp): ACP 4.5 + Alauda OS 标准交付规划(单集群 / 多集群)— 中文审阅稿 Nov 13, 2026
…e 4.5 delivery plan

- State the bootstrap node size (4 cores, 8 GB) and move the link to a note.
- Replace the separate storage cluster with an in-cluster Ceph for workload
  clusters that need storage, and keep DevOps tools with their storage and
  middleware in a dedicated DevOps cluster. Remove virtual machines.
- List ClickHouse alone as the logging choice.
- Split the infra plugin list into one table per cluster layout.
- Point the bootstrap link to its hardware section.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nt notes

- Add review/acp-s2-resource-requirements.zh.md, a review-only breakdown of
  CPU, memory and disk per node pool with the calculation at the end. It sits
  outside docs/, so it is not built or translated, and it will be removed
  before merge.
- State that ClickHouse, Kafka and the registry do not move to infra nodes on
  their own, link the placement settings, and note that the Kafka node label
  cannot change after the instance is created.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… review

- Round CPU to 4/8/12/16/24/32 cores and memory to 8-64 GiB, and pick memory
  so that the node stays under 85 percent.
- Estimate the log components' CPU for 5,000 logs per second from the
  single-node profile instead of summing the three-node profile limits.
- Single-cluster and workload (a) infra: 16 cores / 48 GiB. Global infra:
  12 cores / 32 GiB, with 16 cores / 32 GiB as the uniform-machine option.
- Add an (a) versus (b) breakdown that explains the larger total after
  splitting the non-business nodes into three pools.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lane

Move the platform Ingress Gateway from the global infra pool to the global
control plane. Global infra now needs about 10.8 cores and 22.3 GiB per node,
about 91 percent of 12 cores, and the global control plane stays at 16 cores
and 32 GiB even if the product sizing does not include the gateway.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…very plan

- Replace the machine tables with the reviewed sizes on common machine shapes
  (CPU 4-32 cores, 8-64 GiB, ratios 1:2 to 1:4): infra 16 cores / 48 GiB,
  workload control planes 8 cores / 32 GiB, ingress 8 cores / 16 GiB, log
  12 cores / 48 GiB.
- Add per-node disk tables; round data disks to common sizes (1 TiB TopoLVM
  for log pools, 100 GiB for monitoring).
- Size the in-cluster Ceph as a dedicated storage pool of three 12 cores /
  24 GiB nodes with a 500 GiB OSD disk and a 100 GiB /var/lib/rook disk.
- Apply the same rules in the review file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Relax the machine shape rule to avoid only 1:1 shapes, and restore the
reviewed sizes: global infra 12 cores / 32 GiB (16 cores / 32 GiB as the
uniform option) and workload (b) log 12 cores / 32 GiB. Update the (a)
versus (b) totals in the review file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch was successfully deployed

1 active deployment
translate — 1d7f1b7f Deployed Nov 14, 2026 by chinameok via build #2798
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants