| commit | 3eb1372480bd73436147be8b927665ccffb8cbde | [log] [tgz] |
|---|---|---|
| author | Dmitri Shuralyov <dmitshur@golang.org> | Mon Aug 19 17:42:25 2019 -0400 |
| committer | Dmitri Shuralyov <dmitshur@golang.org> | Tue Aug 20 21:57:39 2019 +0000 |
| tree | c8537f66d12092561e32dce611dc286b90aabe51 | |
| parent | 06ec23726fa04cc83cd329fea55ab0543b7fc6af [diff] |
cmd/gitmirror: replace batched git push mirroring with a single git push In 2016, there were problems with git mirroring being slow and flaky, as described in golang/go#16388. In order to resolve that problem, CL 25110 added a custom mirroring implementation that effectively split up the work done by a single git push -f --mirror dest command into smaller batches of work. It was noted in the commit message and updated documentation of CL 127315 that it may no longer be needed in newer versions of git, and that the problem seemed to affect the HTTPS transport, not so much the SSH transport. By now, the following conditions are different: • It's 2019. • We're using a newer version of git. • We're using the SSH transport (which may or may not be better than the HTTPS transport). • We are mirroring only the refs/heads/* and refs/tags/* namespaces, not refs/changes/* which contains an order of magnitude more refs. As a result, it should be possible to revert to a simple mirroring implementation that uses a single git push -f --mirror dest command. This removes complexity in our code, leveraging the git binary to do more for us (at the expense of relinquishing tighter control and ability to do things differently if we start encountering problems). It resolves the issue of deleted refs not being deleted from mirrors (golang/go#23099) without having to add even more complexity. Debian 10 (buster) is the current stable release of Debian. Start using it, since it comes with a newer version of git (git version 2.20.1, compared to git version 2.11.0 in Debian 9), which we want in hopes of git being better capable of pushing a large number of refs without problems. If we run into problems again, we might have to add some of the complexity back, but hopefully that won't happen. Remove the old and unused code since it's easy to get it back from VCS history. Fixes golang/go#23099 Fixes golang/go#29560 Updates golang/go#16388 Change-Id: Id63419a5e4df5a87d7c9dc4be4e7e38a4849b3df Reviewed-on: https://go-review.googlesource.com/c/build/+/190897 Run-TryBot: Dmitri Shuralyov <dmitshur@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org> Reviewed-by: Alexander Rakoczy <alex@golang.org>
This subrepository holds the source for various packages and tools that support Go's build system and the development of the Go programming language.
This repository uses Gerrit for code changes. To contribute, see https://golang.org/doc/contribute.html.
The main issue tracker for the blog is located at https://github.com/golang/go/issues. Prefix your issue with “x/build/DIR: ” in the subject line.
The main components of the Go build system are:
The dashboard, in app/, serves https://build.golang.org/. It runs on App Engine and holds the state for which builds passed or failed, and stores the build failure logs for post-submit failures. (Trybot build failure logs are stored elsewhere). The dashboard does not execute any builds on its own.
The coordinator, in cmd/coordinator/, serves https://farmer.golang.org/. It runs on GKE and coordinates the whole build system. It finds work to do (both pre-submit “TryBot” work, and post-submit work) and executes builds, allocating machines to run the builds. It is the owner of all machines.
The Go package in buildenv/ contains constants for where the dashboard and coordinator run, for prod, staging, and local development.
The buildlet, in cmd/buildlet/, is the HTTP server that runs on each worker machine to execute builds on the coordinator's behalf. This runs on every possible GOOS/GOARCH value. The buildlet binaries are stored on Google Cloud Storage and fetched per-build, so we can update the buildlet binary independently of the underlying machine images. The buildlet is the most insecure server possible: it has HTTP handlers to read & write arbitrary content to disk, and to execute any file on disk. It also has an SSH tunnel handler. The buildlet must never be exposed to the Internet. The coordinator provisions buildlets in one of three ways:
by creating VMs on Google Compute Engine (GCE) with custom images configured to fetch & run the buildlet on boot, listening on port 80 in a private network.
by running Linux containers (on either Google Kubernetes Engine or GCE with the Container-Optimized OS image), with the container images configured to fetch & run the buildlet on start, also listening on port 80 in a private network.
by taking buildlets out of a pool of connected, dedicated machines. The buildlet can run in either listen mode (as on GCE and GKE) or in reverse mode. In reverse mode, the buildlet connects out to https://farmer.golang.org/ and registers itself with the coordinator. The TCP connection is then logically reversed (using revdial and when the coordinator needs to do a build, it makes HTTP requests to the coordinator over the already-open TCP connection.
These three pools can be viewed at the coordinator's http://farmer.golang.org/#pools
The env/ directory describes build environments. It contains scripts to create VM images, Dockerfiles to create Kubernetes containers, and instructions and tools for dedicated machines.
maintner in maintner/ is a library for slurping all of Go's GitHub and Gerrit state into memory. The daemon maintnerd in maintner/maintnerd/ runs on GKE and serves https://maintner.golang.org/. The daemon watches GitHub and Gerrit and appends to a mutation log whenever it sees new activity. The logs are stored on GCS and served to clients.
The godata package in maintner/godata/ provides a trivial API to let anybody write programs against Go's maintner corpus (all of our GitHub and Gerrit history), live up to the second. It takes a few seconds to load into memory and a few hundred MB of RAM after it downloads the mutation log from the network.
pubsubhelper in cmd/pubsubhelper/ is a dependency of maintnerd. It runs on GKE, is available at https://pubsubhelper.golang.org/, and runs an HTTP server to receive Webhook updates from GitHub on new activity and an SMTP server to receive new activity emails from Gerrit. It then is a pubsub system for maintnerd to subscribe to.
The gitmirror server in cmd/gitmirror/ mirrors Gerrit to GitHub, and also serves a mirror of the Gerrit code to the coordinator for builds, so we don't overwhelm Gerrit and blow our quota.
The Go gopherbot bot logic runs on GKE. The code is in cmd/gopherbot. It depends on maintner via the godata package.
The developer dashboard at https://dev.golang.org/ runs on GKE. Its code is in devapp/. It also depends on maintner via the godata package.
cmd/retrybuilds: a Go client program to delete build results from the dashboard
If you wish to run a Go builder, please email golang-dev@googlegroups.com first. There is documentation at https://golang.org/wiki/DashboardBuilders, but depending on the type of builder, we may want to run it ourselves, after you prepare an environment description (resulting in a VM image) of it. See the env directory.