Gerrit Code Review - System Design

Gerrit is a web based code review system, facilitating online code reviews for projects using the Git version control system.

Gerrit makes reviews easier by showing changes in a side-by-side display, and allowing inline comments to be added by any reviewer.

Gerrit simplifies Git based project maintainership by permitting any authorized user to submit changes to the master Git repository, rather than requiring all approved changes to be merged in by hand by the project maintainer. This functionality enables a more centralized usage of Git.

Google developed Mondrian, a Perforce based code review tool to facilitate peer-review of changes prior to submission to the central code repository. Mondrian is not open source, as it is tied to the use of Perforce and to many Google-only services, such as Bigtable. Google employees have often described how useful Mondrian and its peer-review process is to their day-to-day work.

Guido van Rossum open sourced portions of Mondrian within Rietveld, a similar code review tool running on Google App Engine, but for use with Subversion rather than Perforce. Rietveld is in common use by many open source projects, facilitating their peer reviews much as Mondrian does for Google employees. Unlike Mondrian and the Google Perforce triggers, Rietveld is strictly advisory and does not enforce peer-review prior to submission.

Git is a distributed version control system, wherein each repository is assumed to be owned/maintained by a single user. There are no inherit security controls built into Git, so the ability to read from or write to a repository is controlled entirely by the host’s filesystem access controls. When multiple maintainers collaborate on a single shared repository a high degree of trust is required, as any collaborator with write access can alter the repository.

Gitosis provides tools to secure centralized Git repositories, permitting multiple maintainers to manage the same project at once, by restricting the access to only over a secure network protocol, much like Perforce secures a repository by only permitting access over its network port.

The Android Open Source Project (AOSP) was founded by Google by the open source releasing of the Android operating system. AOSP has selected Git as its primary version control tool. As many of the engineers have a background of working with Mondrian at Google, there is a strong desire to have the same (or better) feature set available for Git and AOSP.

Gerrit Code Review started as a simple set of patches to Rietveld, and was originally built to service AOSP. This quickly turned into a fork as we added access control features that Guido van Rossum did not want to see complicating the Rietveld code base. As the functionality and code were starting to become drastically different, a different name was needed. Gerrit calls back to the original namesake of Rietveld, Gerrit Rietveld, a Dutch architect.

Gerrit 2.x is a complete rewrite of the Gerrit fork, completely changing the implementation from Python on Google App Engine, to Java on a J2EE servlet container and a SQL database.

Developers create one or more changes on their local desktop system, then upload them for review to Gerrit using the standard git push command line program, or any GUI which can invoke git push on behalf of the user. Authentication and data transfer are handled through SSH. Users are authenticated by username and public/private key pair, and all data transfer is protected by the SSH connection and Git’s own data integrity checks.

Each Git commit created on the client desktop system is converted into a unique change record which can be reviewed independently. Change records are stored in PostgreSQL, where they can be queried to present customized user dashboards, enumerating any pending changes.

A summary of each newly uploaded change is automatically emailed to reviewers, so they receive a direct hyperlink to review the change on the web. Reviewer email addresses can be specified on the git push command line, but typically reviewers are automatically selected by Gerrit by identifying users who have change approval permissions in the project.

Reviewers use the web interface to read the side-by-side or unified diff of a change, and insert draft inline comments where appropriate. A draft comment is visible only to the reviewer, until they publish those comments. Published comments are automatically emailed to the change author by Gerrit, and are CC’d to all other reviewers who have already commented on the change.

When publishing comments reviewers are also given the opportunity to score the change, indicating whether they feel the change is ready for inclusion in the project, needs more work, or should be rejected outright. These scores provide direct feedback to Gerrit’s change submit function.

After a change has been scored positively by reviewers, Gerrit enables a submit button on the web interface. Authorized users can push the submit button to have the change enter the project repository. The equivilant in Subversion or Perforce would be that Gerrit is invoking svn commit or p4 submit on behalf of the web user pressing the button. Due to the way Git audit trails are maintained, the user pressing the submit button does not need to be the author of the change.

End-user web browsers make HTTP requests directly to Gerrit’s HTTP server. As nearly all of the user interface is implemented through Google Web Toolkit (GWT), the majority of these requests are transmitting compressed JSON payloads, with all HTML being generated within the browser. Most responses are under 1 KB.

Gerrit’s HTTP server side component is implemented as a standard Java servlet, and thus runs within any J2EE servlet container. Popular choices for deployments would be Tomcat or Jetty, as these are high-quality open-source servlet containers that are readily available for download.

End-user uploads are performed over SSH, so Gerrit’s servlets also start up a background thread to receive SSH connections through an independent SSH port. SSH clients communicate directly with this port, bypassing the HTTP server used by browsers.

Server side data storage for Gerrit is broken down into two different categories:

The Git repository data is the Git object database used to store already submitted revisions, as well as all uploaded (proposed) changes. Gerrit uses the standard Git repository format, and therefore requires direct filesystem access to the repositories. All repository data is stored in the filesystem and accessed through the JGit library. Repository data can be stored on remote servers accessible through NFS or SMB, but the remote directory must be mounted on the Gerrit server as part of the local filesystem namespace. Remote filesystems are likely to perform worse than local ones, due to Git disk IO behavior not being optimized for remote access.

The Gerrit metadata contains a summary of the available changes, all comments (published and drafts), and individual user account information. The metadata is housed in a PostgreSQL database, which can be located either on the same server as Gerrit, or on a different (but nearby) server. Most installations would opt to install both Gerrit and PostgreSQL on the same server, to reduce administration overheads.

User authentication is handled by OpenID, and therefore Gerrit requires that the OpenID provider selected by a user must be online and operating in order to authenticate that user.

Gerrit is developed as a self-hosting open source project:

As a source code review system for open source projects, where the commonly preferred language for communication is typically English, Gerrit does not make internationalization or localization a priority.

The majority of Gerrit’s users will be writing change descriptions and comments in English, and therefore an English user interface is usable by the target user base.

Gerrit uses GWT’s i18n support to externalize all constant strings and messages shown to the user, so that in the future someone who really needed a translated version of the UI could contribute new string files for their locale(s).

Right-to-left (RTL) support is only barely considered within the Gerrit code base. Some portions of the code have tried to take RTL into consideration, while others probably need to be modified before translating the UI to an RTL language.

Whenever possible Gerrit displays raw text rather than image icons, so screen readers should still be able to provide useful information to blind persons accessing Gerrit sites.

Standard HTML hyperlinks are used rather than HTML div or span tags with click listeners. This provides two benefits to the end-user. The first benefit is that screen readers are optimized to locating standard hyperlink anchors and presenting them to the end-user as a navigation action. The second benefit is that users can use the open in new tab/window feature of their browser whenever they choose.

When possible, Gerrit uses the ARIA properties on DOM widgets to provide hints to screen readers.

Supporting non-JavaScript enabled browsers is a non-goal for Gerrit.

As Gerrit is a pure-GWT application with no server side rendering fallbacks, the browser must support modern JavaScript semantics in order to access the Gerrit web application. Dumb clients such as lynx, wget, curl, or even many search engine spiders are not able to access Gerrit content.

As Google Web Toolkit (GWT) is used to generate the browser specific versions of the client-side JavaScript code, Gerrit works on any JavaScript enabled browser which GWT can produce code for. This covers the majority of the popular browsers.

The Gerrit project does not have the development resources necessary to support two parallel UI implementations (GWT based JavaScript and server-side rendering). Consequently only one is implemented.

There are number of web browsers available with full JavaScript support, and nearly every operating system (including any PDA-like mobile phone) comes with one standard. Users who are committed to developing changes for a Gerrit managed project can be expected to be able to run a JavaScript enabled browser, as they also would need to be running Git in order to contribute.

There are a number of open source browsers available, including Firefox and Chromium. Users have some degree of choice in their browser selection, including being able to build and audit their browser from source.

The majority of the content stored within Gerrit is also available through other means, such as gitweb or the git:// protocol. Any existing search engine spider can crawl the server-side HTML produced by gitweb, and thus can index the majority of the changes which might appear in Gerrit. Some engines may even choose to crawl the native version control database, such as ohloh.net does. Therefore the lack of support for most search engine spiders is a non-issue for most Gerrit deployments.

Gerrit integrates with an existing gitweb installation by optionally creating hyperlinks to reference changes on the gitweb server.

Gerrit integrates with an existing git-daemon installation by optionally displaying git:// URLs for users to download a change through the native Git protocol.

Gerrit integrates with any OpenID provider for user authentication, making it easier for users to join a Gerrit site and manage their authentication credentials to it. To make use of Google Accounts as an OpenID provider easier, Gerrit has a shorthand "Sign in with a Google Account" link on its sign-in screen. Gerrit also supports a shorthand sign in link for Yahoo!. Other providers may also be supported more directly in the future.

Site administrators may limit the range of OpenID providers to a subset of "reliable providers". Users may continue to use any OpenID provider to publish comments, but granted privileges are only available to a user if the only entry point to their account is through the defined set of "reliable OpenID providers". This permits site administrators to require HTTPS for OpenID, and to use only large main-stream providers that are trustworthy, or to require users to only use a custom OpenID provider installed alongside Gerrit Code Review.

Gerrit integrates with some types of corporate single-sign-on (SSO) solutions, typically by having the SSO authentication be performed in a reverse proxy web server and then blindly trusting that all incoming connections have been authenticated by that reverse proxy. When configured to use this form of authentication, Gerrit does not integrate with OpenID providers.

When installing Gerrit, administrators may optionally include an HTML header or footer snippet which may include user tracking code, such as that used by Google Analytics. This is a per-instance configuration that must be done by hand, and is not supported out of the box. Other site trackers instead of Google Analytics can be used, as the administrator can supply any HTML/JavaScript they choose.

Gerrit does not integrate with any Google service, or any other services other than those listed above.

Gerrit uses an XSRF protected variant of JSON-RPC 1.1 to communicate between the browser client and the server.

As the protocol is not the GWT-RPC protocol, but is instead a self-describing standard JSON format it is easily implemented by any 3rd party client application, provided the client has a JSON parser and HTTP client library available.

As the entire command set necessary for the standard web browser based UI is exposed through JSON-RPC over HTTP, there are no other data feeds or command interfaces to the server.

Commands requiring user authentication may require the user agent to complete a sign-in cycle through the user’s OpenID provider in order to establish the HTTP cookie Gerrit uses to track user identity. Automating this sign-in process for non-web browser agents is outside of the scope of Gerrit, as each OpenID provider uses its own sign-in sequence. Use of OpenID providers which have difficult to automate interfaces may make it impossible for non-browser agents to be used with the JSON-RPC interface.

Gerrit stores the following information per user account:

The full name and preferred email address fields are shown to any site visitor viewing a page containing a change uploaded by the account owner, or containing a published comment written by the account owner.

Showing the full name and preferred email is approximately the same risk as the From header of an email posted to a public mailing list that maintains archives, and Gerrit treats these fields in much the same way that a mailing list archive might handle them. Users who don’t want to expose this information should either not participate in a Gerrit based online community, or open a new email address dedicated for this use.

As the Gerrit UI data is only available through XSRF protected JSON-RPC calls, "screen-scraping" for email addresses is difficult, but not impossible. It is unlikely a spammer will go through the effort required to code a custom scraping application necessary to cull email addresses from published Gerrit comments. In most cases these same addresses would be more easily obtained from the project’s mailing list archives.

The user’s name and email address is stored unencrypted in the Gerrit metadata store, typically a PostgreSQL database.

The snail-mail mailing address, country, and phone and fax numbers are gathered to help project leads contact the user should there be a legal question regarding any change they have uploaded.

These sensitive fields are immediately encrypted upon receipt with a GnuPG public key, and stored "off site" in another data store, isolated from the main Gerrit change data. Gerrit does not have access to the matching private key, and as such cannot decrypt the information. Therefore these fields are write-once in Gerrit, as not even the account owner can recover the values they previously stored.

It is expected that the address information would only need to be decrypted and revealed with a valid court subpoena, but this is really left to the discretion of the Gerrit site administrator as to when it is reasonable to reveal this information to a 3rd party.

Gerrit makes no attempt to detect spam changes or comments. The somewhat high barrier to entry makes it unlikely that a spammer will target Gerrit.

To upload a change, the client must speak the native Git protocol embedded in SSH, with some custom Gerrit semantics added on top. The client must have their public key already stored in the Gerrit database, which can only be done through the XSRF protected JSON-RPC interface. The level of effort required to construct the necessary tools to upload a well-formatted change that isn’t rejected outright by the Git and Gerrit checksum validations is too high to for a spammer to get any meaningful return.

To post and publish a comment a client must sign in with an OpenID provider and then use the XSRF protected JSON-RPC interface to publish the draft on an existing change record. Again, the level of effort required to implement the Gerrit specific XSRF protections and the JSON-RPC payload format necessary to post a draft and then publish that draft is simply too high for a spammer to bother with.

Both of these assumptions are also based upon the idea that Gerrit will be a lot less popular than blog software, and thus will be running on a lot less websites. Spammers therefore have very little returned benefit for getting over the protocol hurdles.

These assumptions may need to be revisited in the future if any public Gerrit site actually notices spam.

Gerrit targets for sub-250 ms per page request, mostly by using very compact JSON payloads bewteen client and server. However, as most of the serving stack (network, hardware, PostgreSQL metadata database) is out of control of the Gerrit developers, no real guarantees can be made about latency.

Gerrit is designed for a very large scale open source project, or large commerical development project. Roughly this amounts to parameters such as the following:

Out of the box, Gerrit will handle the "Default Maximum". Site administrators may reconfigure their servers by editing gerrit.config to run closer to the estimated maximum if sufficient memory is made avaliable to the JVM and the relevant cache.*.memoryLimit variables are increased from their defaults.

Very few, if any open source projects have more than a handful of Git repositories associated with them. Since Gerrit treats each Git repository as a project, an upper limit of 10,000 projects is reasonable. If a site has more than 1,000 projects, administrators should increase cache.projects.memoryLimit to match.

Almost no open source project has 1,000 contributors over all time, let alone on a daily basis. This default figure of 1,000 was WAG’d by looking at PR statements published by cell phone companies picking up the Android operating system. If all of the stated employees in those PR statements were working on only the open source Android repositories, we might reach the 1,000 estimate listed here. Knowing these companies as being very closed-source minded in the past, it is very unlikely all of their Android engineers will be working on the open source repository, and thus 1,000 is a very high estimate.

The upper maximum of 50,000 contributors is based on existing installations that are already handling quite a bit more than the default maximum of 1,000 contributors. Given how the user data is stored and indexed, supporting 50,000 contributor accounts (or more) is easily possible for a server. If a server has more than 1,000 active contributors, cache.accounts.memoryLimit should be increased by the site administrator, if sufficient RAM is available to the host JVM.

The estimate of 100 changes per day was WAG’d off some estimates originally obtained from Android’s development history. Writing a good change that will be accepted through a peer-review process takes time. The average engineer may need 4-6 hours per change just to write the code and unit tests. Proper design consideration and additional but equally important tasks such as meetings, interviews, training, and eating lunch will often pad the engineer’s day out such that suitable changes are only posted once a day, or once every other day. For reference, the entire Linux kernel has an average of only 79 changes/day. If more than 100 changes are active per day, site administrators should consider increasing the cache.diff.memoryLimit and cache.diff_intraline.memoryLimit.

On average any given change will need to be modified once to address peer review comments before the final revision can be accepted by the project. Executing these revisions also eats into the contributor’s time, and is another factor limiting the number of changes/day accepted by the Gerrit instance. However, even though this implies only 2 revisions/change, many existing Gerrit installations have seen 20 or more revisions/change, when new contributors are learning the project’s style and conventions.

On average, each change will have 2 reviewers, a human and an automated test bed system. Usually this would be the project lead, or someone who is familiar with the code being modified. The time required to comment further reduces the time available for writing one’s own changes. However, existing Gerrit installations have seen 8 or more reviewers frequently show up on changes that impact many functional areas, and therefore it is reasonable to expect 8 or more reviewers to be able to work together on a single change.

Existing installations have successfully processed change reviews with more than 16,000 files per change. However, since 16,000 modified/new files is a massive amount of code to review, it is more typical to see less than 10 files modified in any single change. Changes larger than 10 files are typically merges, for example integrating the latest version of an upstream library, where the reviewer has little to do beyond verifying the project compiles and passes a test suite.

Gerrit’s web UI would require on average 4+F+F*C HTTP requests to review a change and post comments. Here F is the number of files modified by the change, and C is the number of inline comments left by the reviewer per file. The constant 4 accounts for the request to load the reviewer’s dashboard, to load the change detail page, to publish the review comments, and to reload the change detail page after comments are published.

This WAG’d estimate boils down to 216,000 HTTP requests per day (QPD). Assuming these are evenly distributed over an 8 hour work day in a single time zone, we are looking at approximately 7.5 queries per second (QPS).

Gerrit serves most requests in under 60 ms when using the loopback interface and a single processor. On a single CPU system there is sufficient capacity for 16 QPS. A dual processor system should be more than sufficient for a site with the estimated load described above.

Given a more realistic estimate of 79 changes per day (from the Linux kernel) suggests only 8,532 queries per day, and a much lower 0.29 QPS when spread out over an 8 hour work day.

A 24 core server is able to handle ~25 concurrent git fetch operations per second. The issue here is each concurrent operation demands one full core, as the computation is almost entirely server side CPU bound. 25 concurrent operations is known to be sufficient to support hundreds of active developers and 50 automated build servers polling for updates and building every change. (This data was derived from an actual installation’s performance.)

Because of the distributed nature of Git, end-users don’t need to contact the central Gerrit Code Review server very often. For git fetch traffic, slave mode is known to be an effective way to offload traffic from the main server, permitting it to scale to a large user base without needing an excessive number of cores in a single system.

Clients on very slow network connections (for example home office users on VPN over home DSL) may be network bound rather than server side CPU bound, in which case a core may be effectively shared with another user. Possible core sharing due to network bottlenecks generally holds true for network connections running below 10 MiB/sec.

If the server’s own network interface is 1 Gib/sec (Gigabit Ethernet), the system can really only serve about 10 concurrent clients at the 10 MiB/sec speed, no matter how many cores it has.

The average size of a revision in the Linux kernel once compressed by Git is 2,327 bytes, or roughly 2 KiB. Over the course of a year a Gerrit server running with the estimated maxium parameters above might see an introduction of 1.4 GiB over the total set of 10,000 projects hosted in that server. This figure assumes the majority of the content is human written source code, and not large binary blobs such as disk images or media files.

Production Gerrit installations have been tested, and are known to handle Git repositories in the multigigabyte range, storing binary files, ranging in size from a few kilobytes (for example compressed icons) to 800+ megabytes (firmware images, large uncompressed original artwork files). Best practices encourage breaking very large binary files into their Git repositories based on access, to prevent desktop clients from needing to clone unnecessary materials (for example a C developer does not need every 800+ megabyte firmware image created by the product’s quality assurance team).

Gerrit largely assumes that the local filesystem where Git repository data is stored is always available. Important data written to disk is also forced to the platter with an fsync() once it has been fully written. If the local filesystem fails to respond to reads or becomes corrupt, Gerrit has no provisions to fallback or retry and errors will be returned to clients.

Gerrit largely assumes that the metadata PostgreSQL database is online and answering both read and write queries. Query failures immediately result in the operation aborting and errors being returned to the client, with no retry or fallback provisions.

Due to the relatively small scale described above, it is very likely that the Git filesystem and PostgreSQL based metadata database are all housed on the same server that is running Gerrit. If any failure arises in one of these components, it is likely to manifest in the others too. It is also likely that the administrator cannot be bothered to deploy a cluster of load-balanced server hardware, as the scale and expected load does not justify the hardware or management costs.

Most deployments caring about reliability will setup a warm-spare standby system and use a manual fail-over process to switch from the failed system to the warm-spare.

As Git is a distributed version control system, and open source projects tend to have contributors from all over the world, most contributors will be able to tolerate a Gerrit down time of several hours while the administrator is notified, signs on, and brings the warm-spare up. Pending changes are likely to need at least 24 hours of time on the Gerrit site anyway in order to ensure any interested parties around the world have had a chance to comment. This expected lag largely allows for some downtime in a disaster scenario.

PostgreSQL can be configured to save its write-ahead-log (WAL) and ship these logs to other systems, where they are applied to a warm-standby backup in real time. Gerrit instances which care about reduduncy will setup this feature of PostgreSQL to ensure the warm-standby is reasonably current should the master go offline.

Gerrit can be configured to replicate changes made to the local Git repositories over any standard Git transports. This can be configured in '$site_path'/etc/replication.conf to send copies of all changes over SSH to other servers, or to the Amazon S3 blob storage service.

Gerrit does not maintain logs on its own.

Published comments contain a publication date, so users can judge when the comment was posted and decide if it was "recent" or not. Only the timestamp is stored in the database, the IP address of the comment author is not stored.

Changes uploaded over the SSH daemon from git push have the standard Git reflog updated with the date and time that the upload occurred, and the Gerrit account identity of who did the upload. Changes submitted and merged into a branch also update the Git reflog. These logs are available only to the Gerrit site administrator, and they are not replicated through the automatic replication noted earlier. These logs are primarly recorded for an "oh s**t" moment where the administrator has to rewind data. In most installations they are a waste of disk space. Future versions of JGit may allow disabling these logs, and Gerrit may take advantage of that feature to stop writing these logs.

A web server positioned in front of Gerrit (such as a reverse proxy) or the hosting servlet container may record access logs, and these logs may be mined for usage information. This is outside of the scope of Gerrit.

Gerrit is currently manually tested through its web UI.

JGit has a fairly extensive automated unit test suite. Most new changes to JGit are rejected unless corresponding automated unit tests are included.

Reitveld can’t be used as it does not provide the "submit over the web" feature that Gerrit provides for Git.

Gitosis can’t be used as it does not provide any code review features, but it does provide basic access controls.

Email based code review does not scale to a project as large and complex as Android. Most contributors at least need some sort of dashboard to keep track of any pending reviews, and some way to correlate updated revisions back to the comments written on prior revisions of the same logical change.