This directory contains the client-side code for CockroachDB's web-based admin UI, which provides details about a cluster's performance and health. See the Admin UI docs for an expanded overview.
To start developing the UI, be sure you're able to build and run a CockroachDB
node. Instructions for this are located in the top-level README. Every Cockroach
node serves the UI, by default on port 8080, but you can customize the port with
the --http-port
flag. If you've started a node with the default options,
you'll be able to access the UI at http://localhost:8080.
Our UI is compiled using a collection of tools that depends on Node.js and are managed with Yarn, a package manager that offers more deterministic package installation than NPM. NodeJS 6.x and Yarn 1.7.0 are known to work. Chrome, Google's internet browser. Unit tests are run using Chrome's "Headless" mode.
With Node and Yarn installed, bootstrap local development by running make
in
this directory. This will run yarn install
to install our Node dependencies,
run the tests, and compile the assets. Asset compilation happens in two steps.
First, Webpack runs the TypeScript compiler and CSS
preprocessor to assemble assets into the dist
directory. Then, we package
those assets into embedded.go
using
go-bindata. When you later run make build
in the parent directory, embedded.go
is linked into the cockroach
binary so that it can serve the admin UI when you run cockroach start
.
When making changes to the UI, it is desirable to see those changes with data from an existing cluster without rebuilding and relaunching the cluster for each change. This is useful for rapidly visualizing local development changes against a consistent and realistic dataset.
We've created a simple NodeJS proxy to accomplish this. This server serves all requests for web resources (JavaScript, HTML, CSS) out of the code in this directory, while proxying all API requests to the specified CockroachDB node.
To use this proxy, in Cockroach's root directory run:
$ make ui-watch TARGET=<target-cluster-http-uri>
or, in pkg/ui
run:
$ make watch TARGET=<target-cluster-http-uri>
then navigate to http://localhost:3000
to access the UI.
To proxy to a cluster started up in secure mode, in Cockroach's root directory run:
$ make ui-watch-secure TARGET=<target-cluster-https-uri>
or, in pkg/ui
run:
$ make watch-secure TARGET=<target-cluster-https-uri>
then navigate to https://localhost:3000
to access the UI.
While the proxy is running, any changes you make in the src
directory will
trigger an automatic recompilation of the UI. This recompilation should be much
faster than a cold compile—usually less than one second—as Webpack can reuse
in-memory compilation artifacts from the last compile.
If you get cryptic TypeScript compile/lint failures upon running make
that
seem completely unrelated to your changes, try removing yarn.installed
and
node_modules
before re-running make
(do NOT run yarn install
directly).
Be sure to also commit modifications resulting from dependency changes, like
updates to package.json
and yarn.lock
.
To improve Webpack compile times, we split the compile output into three bundles, each of which can be compiled independently. The feature that enables this is Webpack's DLLPlugin, named after the Windows term for shared libraries ("dynamic-link libraries").
Third-party dependencies, which change infrequently, are contained in the vendor DLL. Generated protobuf definitions, which change more frequently, are contained in the protos DLL. First-party JavaScript and TypeScript are compiled in the main app bundle, which is then "linked" against the two DLLs. This means that updating a dependency or protobuf only requires rebuilding the appropriate DLL and the main app bundle, and updating a UI source file doesn't require rebuilding the DLLs at all. When DLLs were introduced, the time required to start the proxy was reduced from over a minute to under five seconds.
DLLs are not without costs. Notably, the development proxy cannot determine when
a DLL is out-of-date, so the proxy must be manually restarted when dependencies
or protobufs change. (The Make build system, however, tracks the DLLs'
dependencies properly, so a top-level make build
will rebuild exactly the
necessary DLLs.) DLLs also make the Webpack configuration rather complicated.
Still, the tradeoff seems well worth it.
In CCL builds, code in pkg/ui/ccl/src
overrides code in pkg/ui/src
at build
time, via a Webpack import resolution rule. E.g. if a file imports
src/views/shared/components/licenseType
, it'll resolve to
pkg/ui/src/views/shared/components/licenseType
in an OSS build, and
pkg/ui/ccl/src/views/shared/components/licenseType
in a CCL build.
CCL code can import OSS code by prefixing paths with oss/
, e.g.
import "oss/src/myComponent"
. By convention, this is only done by a CCL file
importing the OSS version of itself, e.g. to render the OSS version of itself
when the trial period has expired.
To run the tests outside of CI:
$ make test
The regular build also produces a webpage with a report on the bundle size. Build the app, then take a look with:
$ make build
$ open pkg/ui/dist/stats.ccl.html
Or, to view the OSS bundle:
$ make buildoss
$ open pkg/ui/dist/stats.oss.html
To comply with the SIL Open Font License, we have reproducible builds of our WOFF font bundles based on the original TTF files sourced from Google Fonts.
To rebuild the font bundles (perhaps to bring in an updated version of a typeface),
simply run make fonts
in the UI directory (or make ui-fonts
elsewhere). This
requires fontforge
to be available on your system. Then validate the updated
fonts and commit them.
To add a new typeface, edit the script scripts/font-gen
to fetch and convert it,
and then add it to styl/base/typography.styl
to pull it into the bundle.
The NPM registry (and the Yarn proxy in front of it) have historically proven
quite flaky. Errors during yarn install
were the leading cause of spurious CI
failures in the first half of 2018.
The solution is to check our JavaScript dependencies into version control, like we do for our Go dependencies. Checking in the entire node_modules folder is a non-starter: it clocks in at 230MB and 28k files at the time of writing. Instead, we use a Yarn feature called the offline mirror. We ship a .yarnrc file that instructs Yarn to save a tarball of each package we depend on in the yarn-vendor folder. These tarballs are then checked in to version control. To avoid cluttering the main repository, we've made the yarn-vendor folder a Git submodule that points at the cockroachdb/yarn-vendored repository.
Let's pretend you want to add a dependency on left-pad
. Just use yarn add
like you normally would:
$ cd $GOPATH/src/github.com/cockroachdb/cockroach/pkg/ui
$ yarn add FOO
When Yarn finishes, git status
will report something like this:
$ git status
Changes not staged for commit:
modified: pkg/ui/package.json
modified: pkg/ui/yarn-vendor (untracked content)
modified: pkg/ui/yarn.lock
The changes to package.json and yarn.lock are the normal additions of the new dependency information to the manifest files. The changes to yarn-vendor are unique to the offline mirror mode. Let's look more closely:
$ git -C yarn-vendor status
Untracked files:
left-pad-1.3.0.tgz
Yarn has left you a tarball of the new dependency. Perfect! If you were adding a more complicated dependency, you'd likely see some transitive dependencies appear as well.
The process from here is exactly the same as updating any of our other vendor submodules. Make a new branch in the submodule, commit the new tarball, and push it:
$ cd yarn-vendor
$ git checkout -b YOURNAME/add-left-pad
$ git add .
$ git commit -m 'Add [email protected]'
$ git push origin add-left-pad
Be sure to push to cockroachdb/yarn-vendored directly instead of a personal fork. Otherwise TeamCity won't be able to find the commit.
Then, return to the main repository and commit the changes, including the new submodule commit. Push that and make a PR:
$ cd ..
$ git checkout -b add-left-pad
$ git add pkg/ui
$ git commit -m 'ui: use very smart approach to pad numbers with zeros'
$ git push YOUR-REMOTE add-left-pad
This time, be sure to push to your personal fork. Topic branches are not permitted in the main repository.
When your PR has been approved, please be sure to merge your change to yarn-vendored to master and delete your topic branch:
$ cd yarn-vendor
$ git checkout master
$ git merge add-left-pad
$ git push origin master
$ git push origin -d add-left-pad
This last step is extremely important! Any commit in yarn-vendored that is
referenced by the main repository must remain forever accessible, or it will be
impossible for future git clone
s to build that version of Cockroach. GitHub
will garbage collect commits that are not accessible from any branch or tag, and
periodically, someone comes along and cleans up old topic branches in
yarn-vendored, potentially removing the only reference to a commit. By merging
your commit on master, you ensure that your commit will not go missing.
Our build system is careful to invoke yarn install --offline
, which instructs
Yarn to exit with an error if it would need to reach out to the network. Running
CI on your PR is thus sufficient to verify that all dependencies have been
vendored correctly.
You can perform the verification locally if you'd like, though:
$ cd $GOPATH/src/github.com/cockroachdb/cockroach/pkg/ui
$ rm -r node_modules yarn.installed
$ yarn cache clean
$ make
If make
succeeds, you've vendored dependencies correctly.
To remove a dependency, just run yarn remove
.
Note that removing a dependency will not remove its tarball from the yarn-vendor folder. This is not as bad as it sounds. When Git fetches submodules, it always performs a full clone, so it would wind up downloading deleted tarballs anyway when it fetched older commits.
TODO(benesch): Yarn's offline mode has an additional option,
yarn-offline-mirror-pruning
, that cleans up removed dependencies' tarballs.
Look into using this once dcodeIO/protobuf.js#716 is resolved. At the moment,
ProtobufJS tries to install some dependencies at runtime by invoking npm install
itself (!); we avoid this by running yarn install
on its behalf, but
this means two separate packages share the same yarn-vendor folder and this
confuses Yarn's pruning logic.
If the size of the yarn-vendored repository becomes problematic, we can look into offloading the large files into something like Git LFS. This is contingent upon resolving the above TODO.