Skip to content

chore:bump go to 1.27 - #25

Open
Amazon2025 wants to merge 33 commits into
kite-plus:mainfrom
Amazon2025:main
Open

Amazon2025 wants to merge 33 commits into
kite-plus:mainfrom
Amazon2025:main

Conversation

@Amazon2025

Copy link
Copy Markdown

Go 1.27.4 compiles the release binaries. The go line moves to 1.27.0 so the
toolchain line stays a pin above it, rather than an equal duplicate that
go mod tidy drops.

golangci-lint moves to v2.14.0 with it: v2.13.2 bundles x/tools v0.49.0, which
reads export data up to version 4, and go1.27.2 writes version 5.
Also go 1.27 replaced the compress/flate encoder, so the same image deflates to
different bytes. The rendering version is part of every cache key and the
test pins the output hashes, so both move together. Only PNG output changed;
jpeg, webp and gif reproduce as before.

Go 1.27.4 compiles the release binaries. The go line moves to 1.27.0 so the
toolchain line stays a pin above it, rather than an equal duplicate that
go mod tidy drops.

golangci-lint moves to v2.14.0 with it: v2.13.2 bundles x/tools v0.49.0, which
reads export data up to version 4, and go1.27.2 writes version 5.
Go 1.27 replaced the compress/flate encoder, so the same image deflates to
different bytes. The rendering version is part of every cache key and the
test pins the output hashes, so both move together. Only PNG output changed;
jpeg, webp and gif reproduce as before.
The READMEs, the reference manuals and the architecture notes still asked for
Go 1.26.
VS Code writes .vscode/settings.json per machine and it is not part of the
project.
The test step of make check fails intermittently on Windows with a TempDir
RemoveAll error. The assertion has already run by then; what fails is removing
a directory the system still has a handle on.
Go 1.27.4 compiles the release binaries. The go line moves to 1.27.0 so the
toolchain line stays a pin above it, rather than an equal duplicate that
go mod tidy drops.

golangci-lint moves to v2.14.0 with it: v2.13.2 bundles x/tools v0.49.0, which
reads export data up to version 4, and go1.27.2 writes version 5.
Also go 1.27 replaced the compress/flate encoder, so the same image deflates to
different bytes. The rendering version is part of every cache key and the
test pins the output hashes, so both move together. Only PNG output changed;
jpeg, webp and gif reproduce as before.
v2.13.2 ships x/tools v0.49, whose importer stops at export data version 4;
go 1.27.2 moved the format to version 5, so typechecking the standard library
failed with "export data version 5 is greater than maximum supported version
4". v2.14.0 carries x/tools v0.50 and reads it.

Downloading the binary keeps the job fast, and the pin matches the Makefile's
LINT_VERSION.
@amigoer

amigoer commented Nov 14, 2026

Copy link
Copy Markdown
Contributor

Thanks for this. The bump itself checks out: on go1.27.4, go test ./..., go vet and gofmt all pass locally, and the new PNG hash reproduces.

One thing blocks it: the Dockerfile still has ARG GO_VERSION=1.26.8 ("Must match the toolchain directive in go.mod"). The golang images set GOTOOLCHAIN=local, so with go 1.27.0 in go.mod the build stage stops at go mod download:

go: go.mod requires go >= 1.27.0 (running go 1.26.8; GOTOOLCHAIN=local)

That breaks make docker and the release workflow's image job, and PR CI doesn't build the image, so it stays green. Your first commit had ARG GO_VERSION=1.27.4, but that line got lost when the commits were redone. golang:1.27.4-alpine is published, so restoring it should be all that's needed.

Two small things while you're there:

  • /.vscode in .gitignore isn't part of the Go bump. A global gitignore (core.excludesFile) keeps local editor files out without changing the repo.
  • The title needs a space after the colon: chore: bump go to 1.27.

Video was on no stage. The media library is the only place it fits: it already
owns the upload entrance, which is where the 32 MB in-memory limit has to be
replaced per type and streamed to disk, and B01 already gives one article's
attachments a list, replace, delete and reference check.

The editor's video node has no stage of its own either, so it rides along.
A body could already call one, but the default theme had no template for it,
so a video was the one thing an author could not publish without a theme of
their own.

The syntax stays Hugo's, {{< video src="clip.mp4" >}}: a private !video[]()
would cost a markdown parser, be invisible to a site moving to another
generator, and sit beside shortcodes doing the same job.

src is required; poster, caption, width, loop, muted, captions and lang are
not. The type follows the extension, so webm needs no extra parameter, and the
width is a hint the theme's own max-width rule overrides on a narrow screen.
The default theme ships the template and the style rule, so a body that calls a
video shortcode renders instead of stopping the build. The checklist row says
what it covers rather than asking for it.
The dashboard had no way to tell how much a site holds. GET /storage answers
with the bytes and the number of files under content/, measured by the site
because where a project keeps its files is the store's business.

The walk stops as soon as the request does, and a content directory that does
not exist yet is nothing rather than an error.
The checklist item and the placeholder table both said the dashboard had no
way to say how much a site holds; they say what now answers it.
The footer kept the version in a tooltip, so a badge pointing at a newer
release had nothing to sit beside. The version is on the page now, and when
/site reports a newer release the badge links to its notes.

No request is made here: the server already asks, and the answer rides along
with the site description.
The lint job runs misspell with the US locale, so the British form fails it.
make check runs a linter, and it finds what go vet and gofmt do not: a
misspelled word, a symbol nothing uses, an error that is not wrapped. The
manuals said make check was the gate without saying what it holds.
The admin's assets go out uncompressed, and the largest script is 564 kB where
gzip would make it 182 kB. Whether Kite compresses or leaves it to a proxy
belongs with the remote-access design, so the question is recorded there.
version said what this build is. --check also says whether a newer release is
out, reusing the answer the studio already keeps in the user's cache for a day.

It says plainly when checks are turned off or GitHub cannot be reached, rather
than pretending the build is current.
The manuals explained when a running server asks for a newer release, but not
what a person can ask for themselves. They also said a build from source never
asks, which is only true when nobody tells it to.
The section had Kite asking, saying and telling. It now makes requests,
reports and records, which is what the code does.
Kite, the studio, the server, the build and the workflow said, asked and told.
They now report, state, notify and query: the same facts, in terms the code
works in.

Left alone are the places where a file or a template carries a value, as in
"kite.yaml says" or "T says 8 posts", which name what the data holds rather
than what the program does.
A tag was stamped as v0.1.12-30-g5208135; the leading v is dropped so a build
made here reads like a released one, 0.1.12 at the tag and 0.1.12-30-g5208135
after it.

A binary linked without those flags reported none and unknown. It now takes the
revision and date from the VCS information the toolchain embeds, and --short
prints the version alone for a script to read.
- 被劈成互不相容的两半 → 分成两类互不通用的产品
- 把两件本应正交的事耦死了 → 把两件本应正交的事绑成了一件
- 会直接干掉被监听的 inode → 会直接让被监听的 inode 失效
- 写两套再抹平 / 任何一个破掉 → 写两套再对齐 / 任何一个出问题

只改动词,程度词与原意保留。
- 撑爆队列 → 超出队列容量
- 硬塞进 V1 会拖垮进度 → 一并放进 V1 会拖慢进度
- 这个模型崩掉 → 这个模型不再成立

全库 29 份 md 扫描后只剩这三处属明显口语动词;引号内的用户原话
(「这个 CMS 把我的文件搞乱了」)按原文保留。
按 Euchien 的判断:
- 720 行 把 URL 塞进 .Page 字段 → 写进(措辞更准确)
- 824 行 模型崩掉 → 保持原样(属正常表达,撤回上一次改动)
- architecture 226 行「塞进 Status 字段」不改(中文技术写作惯用)
The minimum was 1.27.0 with a toolchain line above it, which let a build at
1.27.0 through: that release predates the security fixes of 2026-11-10.

The go line now names 1.27.4 itself, so an older toolchain is refused rather
than used. The release pin is read from that line instead of the toolchain
line, which an equal go line would make go mod tidy drop.
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