r-hub: the everything-builder the R community needs
Gábor Csárdi, based on previous versions by J.J. Allaire (RStudio), Ben Bolker (McMaster University) Dirk Eddelbuettel (Debian), Jay Emerson (Yale University), Nicholas Lewin-Koh (Genentech), Joseph Rickert (Revolution), David Smith (Revolution), Murray Stokely (Google), and Simon Urbanek (AT&T)
July 21, 2015
The infrastructure available for developing, building, testing, and validating R packages is of critical importance to the R community. CRAN and R-Forge have traditionally met these needs, however the maintenance and enhancement of R-Forge has significant costs in both money and time. This proposal outlines r-hub, a service that is complementary to CRAN and R-Forge, that would add capabilities, improve extensibility, and create a platform for community contributions to r-hub itself.
- Services that ease all steps the R package development process, creating a package, building binaries and continuous integration, publishing, distributing and maintaining it.
- Make these services free for all members of the community.
- Allow community contributions to r-hub itself.
- Make CRAN maintainers' work easier by pre-testing CRAN package submissions.
- Reuse as much existing technology as possible, especially the tools and services already widely used among R package developers.
- Keep compatibility with current systems.
- Cover as much of the user base as possible.
- Create open systems: provide APIs, make all code for the service open source, accept contributions to it. Make all software open source on GitHub. GitHub can be used with git and svn natively, and also with Mercurial via a plugin.
- Automated system. Make the daily operation of r-hub fully automatic, without any human intervention needed.
- Cost-effective services. Rent machines from the cloud, if possible. Adapt number and type of machines to the current workload.
- Make it possible for organizations and individuals to contribute (the cost of) machines to the build farm.
- Make services reproducible. It should be possible to start up an instance of any r-hub service within minutes, for anyone, even on their own laptop. This is crucial for development, testing and user contribution.
- Separation is key, so use containers and reproducible VM images as much as possible.
To check and build R packages. It also creates binaries for various platforms, including the currently supported CRAN platforms, and potentially others:
- MS Windows
- OSX, each flavor supported by CRAN
- Linux flavors
Users can upload their packages through a web page, or an HTTP API, potentially select platforms to build on. They are provided with a randomized URL where they can check the status of their builds, see their output, and they can also cancel the build. Once the builds are finished they can download source and binary R packages.
Continuous integration for R packages
Make the build server work as a continuous integration service, free to use for all R community members, integrated into GitHub and potentially other popular source code repositories. It will work very similarly to Travis or other GitHub integrated CI services, but will specialize in R packages: better start-up time using caching and binary builds of popular or all packages, no or minimal configuration requirements for the users.
Distribution of R package sources and binaries
Store built source and binary packages in a repository from which they can be installed easily. This service can be used for distributing development versions of CRAN R packages, or for R packages that are not (yet) fit to be distributed on CRAN.
Community web site
To showcase the packages, and to integrate all services. Users can browse packages, search for packages, search the source code, view build statistics, link to GitHub issues.
Implementation details and time line
Build server: Jenkins in a container (week 1)
Jenkins server, running in a Docker container. We need one master server to start with. We will run it in the cloud, initially on DigitalOcean or AWS. We will use dokku to manage it (and will also manage other microservices similarly in containers). We will probably switch to something more integrated in the future, like Kubernetes on CoreOS, but initial development dokku is satisfactory.
We will not expose Jenkins directly, but both users and admins will access it through an HTTP API (the r-hub API). The reason for this is that we might need to use something else than Jenkins for Solaris, and that in the future we might need multiple Jenkins servers.
Initially the server providing the r-hub API is very simple, and will essentially serve as a proxy to our single Jenkins master.
The actual build submissions to Jenkins will be handled by another process, so that the web app can handle requests promptly. The web app simply puts down the file into a shared folder (putting a random submission id in the file name to make sure it is unique), and then adds the a request including the file name and and the id to a RabbitMQ queue. Another process picks up jobs from the queue, and then communicates with Jenkins to add a job and start the build. Query parameters are also passed to the queue.
r-hub API (weeks 1-2)
An HTTP API to manipulate Jenkins and the builds. The protocol is in JSON. Initial design is very simple, this is sufficient for the web-based submissions.
Submit a source R package to be checked and built. The server responds with a JSON string including a random ID that can be used to query the submission later. Query parameters can be used later to select build OS, email notification, etc.
Query the status of a submission. The server responds with a status code (e.g. SUBMITTED, QUEUED, DONE, etc), and if the build is already generating output, a summary of the output as well. The response also contains info about the build machine, if that is known already.
Query the console output of the check.
Worker machines, API for administrators
Query the list of active worker machines.
Instruct Jenkins to create another worker. It's parameters are supplied in the body of the request, in JSON.
Remove a worker.
Jenkins Linux workers (weeks 2-3), backends
We will support different backends for workers. Initially there will be two backends: local, and either DigitalOcean or AWS.
The local backend is for development and testing, it creates a worker container on the local machine.
The DigitalOcean/AWS worker is mainly for production (but can be used for testing). DigitalOcean has an API to create a droplet with Docker support, AWS has an API as well. Jenkins has a Docker plugin that can create the container, once a Docker environment is set up.
Each platform needs to be able to install a recent version of R-devel (and R-patched). For the Linux workers, a permanent Jenkins job will build R-devel daily, in a separate container, and then distribute the (successful) builds via HTTP. The worker containers download the binary R-devel versions upon start-up, and they also check if they have the most recent version at the beginning of each build. This is a simple and cheap HTTP HEAD query.
Current Linux platforms on CRAN:
Other popular Linux platforms we want to support:
r-develon most recent Ubuntu LTS
r-releaseon most recent Ubuntu LTS
r-develon stable Debian
r-releaseon stable Debian
System requirements (week 4)
Create a database of system requirements, in the spirit of https://github.com/metacran/sysreqs The workers need to use this database.
Repositories needed so far:
rhub-jenkinsThe Docker config for our Jenkins instance. It contains the Jenkins config as well.
rhub-appThe web app to submit and query jobs, in node.js.
rhub-apiWeb pages that document the API.
rhub-queueThe app that handles the queue, picks up the files, creates Jenkins jobs, and starts the build.
rhub-worker-*One repo for the Docker config for each worker type. I.e. one for
rhub-backend-localThe local backend. All the code needed to create a local worker, and also the code needed to destroy it. In node.js, probably, so that the app can easily use it.
rhub-backend-doDigitalOcean backend to create/destroy workers on DO.
rhub-backend-awsBackend to create/destroy workers on AWS.
sysreqsDatabase of system requirement mappings.
sysreqs-appWeb app with an API for system requirements.
rhub-issuesA repository for user feedback.
rhub-configEncrypted r-hub configuration, accessible only via the private keys of the admins. All sensitive information needed for r-hub to work will be included here.
Online submission system (week 5)
Test the Linux builders and the system requirements by building all CRAN packages (that are available on Linux). Fix errors, 99% of the CRAN packages are expected to build correctly.
In general, orchestrate, document, test and make public the web app for submitting builds for Linux machines. At this point, we will have the Linux-builder equivalent of Win-builder, as an open, extensible system.
Redundant services, logging and monitoring (weeks 6-7)
For logging, redirect all logs through the network to a single log server. These logs do not include console output from builds, those are handled separately. This is only about the r-hub service logs.
For build logs, at this point of the project, we will just leave them in Jenkins for a while, and then remove them (and the whole job) after a 3-5 days period.
Create a sensu dashboard and the required reporters, to oversee all services.
Redundancy: workers are disposable, so we don't need to worry about them. The web app can be replicated on a second server and then load balanced using AWS services. The same applies to the queue manager.
rhub-monitorConfig for running the monitoring server.
rhub-loggerConfig for running the logger.
rhub-dashboardWeb-app to view the actual monitoring dashboard.
Windows workers (weeks 8-9)
Users will be able to select windows in the submission web app, and builder selection will be part of the r-hub API as well. At this point the user will have to select a single builder.
We can run windows VMs on AWS. We will use code from the r-appveyor project to build the windows worker backend. This project builds a disk image containing a binary R-devel build daily.
rhub-backend-aws-windowsAWS windows backend.
OSX workers (weeks 10-11)
For legal reasons, OSX workers need to run on Apple hardware. Many companies offer OSX as IaaS, and this seems to be the simplest and cheapest solution for us. Most likely we will rent one or two Mac Mini servers with 16GB memory from https://macstadium.com, unless we get a better offer. This server comes with a VMware ESXi v6.0 Hypervisor, so we run multiple instances of OSX on it, in virtual machines. We create snapshot images for the latest two or three OSX versions, and we start each R package build from a snapshot image, assuming that start-up takes a reasonable time (1-2 minutes maximum). If it takes longer, then we will look into running the build as a restricted user, clean up after the build, reuse the VM for multiple builds, and only reboot/recreate the VM once a day.
Jenkins supports ESXi via a plugin, or we can just use the native ESXi API directly from our web-app middleware, to manage the VMs.
For OSX, a daily build of R-devel is available from r-project.org and the workers just download this and install it to a read-only shared area, so that the worker images do not need to be rebuilt daily.
On-demand workers (week 12)
While we strive to have a quick build start-up time, we want to make sure that workers are not idle for long. For OSX at https://macstadium.com, flexible scaling is not available for the rented Mac Mini servers. For Linux and Windows workers, they are typically billed by the hour. We will work out a simple heuristics to shut down idle Windows and Linux workers. This could work from within Jenkins or the r-hub application.
CRAN presubmission (starting from week 13)
We will work with CRAN on the integration of r-hub into the CRAN submission process. At present, 80% of the CRAN submissions are turned down. We set out to reduce this to 20%, by providing feedback to package developers, prior to submitting to CRAN.
The CRAN submission process will have the following steps:
- User submits through web-app on r-hub.
- Package is built and checked on Linux with r-devel. If there are warnings or errors, the submission is rejected.
- Package is built and checked by r-hub, on all CRAN platforms. If there are warnings or errors, the package is rejected.
- The submission is sent to CRAN, including all checks. New warnings and errors are marked for CRAN maintainers. They have the possibility to accepts the submission, and then the source package moves to CRAN's system. Otherwise CRAN maintainers can send back an email to the package maintainer, including all build logs, and possibly their own annotation.
CI for GitHub projects (weeks 14-15)
An R specific CI service, with GitHub integration. It should be completely automatic, with some sane defaults, e.g. by default build and check with r-devel, on Linux (Debian), Windows and the most recent OSX platform. Other platforms can be selected via configuration, or as one-time builds from the web app or the API.
Having a CI also means that we will have projects, and need a database to store their settings, build history, etc. We will use a MongoDB database for this, except for the build output, which will be compressed and stored in a key-value datastore, e.g. Redis or something more lightweight. We will purge the output of old builds if we need to lower the storage costs.
rhub-ciWeb app that handles the hooks from GitHub (and possibly other services in the future).
rhub-ci-docsDocumentation of the r-hub CI.
Binary and source packages from GitHub will be stored on a storage server in a CRAN-like repository, and we will open this repository for public use.
Design a system to handle dependencies among packages stored at CRAN, r-hub, BioConductor, GitHub, etc. This should be part of the DESCRIPTION file, and handled by the r-hub CI automatically.
Open, documented API (week 16-17)
Document and make public the r-hub CI service, including an API to add/remove jobs, run/query builds. Stabilize services. Make sure that almost all CRAN packages build on all three major platforms.
rhubR package, to query the r-hub API: submit packages to build, search packages, etc.
Community website (weeks 18)
Simple web-site, where users can browse and search packages, both in the r-hub repository and CRAN. It will be integrated into the already existing web pages. We will possibly integrate this with the existing http://www.r-pkg.org web pages.
Reverse dependency checks (week 19)
We will include the check of reverse dependencies in CRAN pre-submissions. We compare the new check results to old ones, and include the differences (i.e. newly appearing warnings and errors) in our email notification to CRAN, if a package maintainer decides to continue with the submission, even if reverse dependencies break.
Solaris builds (week 20)
We will look into having Solaris workers. This is really only important for packages with compiled code. It is also non-trivial, for two reasons. First, Jenkins does not officially support Solaris since late 2014. Second, ideally we would need Solaris Sparc as well, as it has a different byte-order than all other CRAN platforms, and many errors only come up there.
We can run Solaris x86 on AWS in a VM, and we can rent a Solaris Sparc LPAR in the cloud, and we can use Solaris Zones containers on it.
Stabilize, publicize r-hub (weeks 21-24)
These four weeks also serve as reserves, if any of the sub-projects run out of time, which is not unlikely.
Some facts first, we base our cost estimates on these:
- There are about 7,000 packages on CRAN, and we estimate that this number will double in five years, to about 14000 packages.
- About 30 packages are updated on CRAN every day, we estimate that this number will grow to 60 in the next five years.
- A CRAN package has on average 2.65 hard (Depends, Imports) dependencies.
- There are about 100,000 repositories on GitHub that are classified as written mainly in R. Out of them about 2,000 are currently using Travis CI.
- On a modern processor (e.g. Xeon 2.9GHz) it takes on average 30 seconds to build and check a CRAN package on Linux, and about twice as long on OSX and three to five times as long on Windows. These numbers assume that binary packages are available, and no container start-up time.
Our estimates for the number of daily builds and required CPU time one year after the start of the project, assuming r-hub is integrated into the CRAN submission process and it is popular on GitHub as a CI service:
- 35 successful CRAN submissions, and 200 unsuccessful ones, a total number of 235 submissions, triggering 235 direct Linux builds on Debian, and about 50 builds on other Linux flavors and other OSes. This also triggers about 150 reverse dependency checks on the Debian Linux.
- We estimate 1,000 CI builds from GitHub, although this number might be highly inaccurate, as we have essentially no real data on this. (The mean number of ~5 pushes per R GitHub repository is useless, as most of these repositories are essentially inactive.) The 1,000 CI builds will be done on Debian Linux, Windows and the most recent OSX platform.
- This is a total of 1,385 builds on the primary Linux platform, 1,050 builds on the primary Windows and primary OSX platforms, and 50 builds on other platforms.
- Assuming 2 minutes container start-up time, this amounts to about 60 CPU hours for the primary Linux platform, 50 CPU hours for the primary OSX platform and 70 CPU hours for the primary windows platform. For other platforms the required resource are negligible.
These estimates are the lower bounds for running a useful service. They do not contain builds that are stuck and until the end of the build time limit (20-40 minutes), and they do not contain CPU time required to build R-devel daily, etc. We estimate that these are negligible compared to the total cost.
Cost for development work:
- $100 per hour, total of 1000 hours over 9-12 months. $100,000 total.
- Domain name registration, virus scanner for the Windows builder, SSL certificates, maximum $500 total.
Ongoing operational costs, from month one:
- Development and production environments, Jenkins server and web server hosting, all server side applications, about $100 per month.
From month two:
- Linux build servers: $100 per month, $10 per platform.
- OSX build servers: one mac mini server $120 per month.
- Windows servers: four t2.medium instances on AWS, $210 per month.
- Solaris server: Solaris Sparc $300 per month, Solaris x86 $40 per month.
- Storage server for binary and source packages: under $200 per month.
Maintenance cost after the first year:
- Ten hours work per month, for one person, $1000 per month.
Total monthly operational cost after the first year: $2112 per month.
They may be prioritized to be included in the project if there is time available from the allocated 1000 hours, and the community needs them.
- Build native binary packages for popular Linux distribution, e.g. Debian, Ubuntu, RedHat Linux or CentOS.
- Builds with memory sanitizers and debuggers: valgrind, UB-SAN, etc.
- Digitally signed binaries, served over HTTPS.
- R CI: since we build R-devel regularly anyway, we can also run tests for it, and multi-platform continuous integration.
- Separate r-hub-devel and r-hub-stable repositories.
- Automatic vertical scaling of build workers. Keep track of how much memory they need and use bigger machines only if needed.
- Check CRAN packages with all allowed major R versions. E.g. if a package depends on R (>= 3.0.0), then check it with R 3.0.3, 3.1.3, 3.2.1.
Community web site and r-hub web app
- RSS feeds for new and updated packages, similarly to CRANberries, with filters for keywords, dependencies, authors.
- Searchable documentation of R packages, online in HTML, both for CRAN and r-hub packages. With inline plots.
- Explain R CMD check failures. Link to relevant parts of the documentation.
- Badges for package versions, number of installs, downloads, checks.
- RSS or email notifications for CRAN check failures.