opendev/gerrit
Code Issues Proposed changes

Merge "Update designdoc to current state"

Browse Source
This commit is contained in:
Han-Wen Nienhuys
2021-03-11 16:13:44 +00:00
committed by Gerrit Code Review
parent f253fa78be d7873e6071
commit ae8b54b741
1 changed files with 302 additions and 401 deletions
Download Patch File Download Diff File Expand all files Collapse all files

703
Documentation/dev-design.txt
View File

@@ -18,83 +18,45 @@ centralized usage of Git.
== Background
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
inherent 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.
filesystem or network access controls.
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 objective of Gerrit is to facilitate Git development by larger
teams: it provides a means to enforce organizational policies around
code submissions, eg. "all code must be reviewed by another
developer", "all code shall pass tests". It achieves this by
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 an SQL database.
Since Gerrit 3.x link:note-db.html[NoteDb] replaced the SQL database
and all metadata is now stored in Git.
* link:http://video.google.com/videoplay?docid=-8502904076440714866[Mondrian Code Review On The Web,role=external,window=_blank]
* link:https://github.com/rietveld-codereview/rietveld[Rietveld - Code Review for Subversion,role=external,window=_blank]
* link:http://eagain.net/gitweb/?p=gitosis.git;a=blob;f=README.rst;hb=HEAD[Gitosis README,role=external,window=_blank]
* link:http://source.android.com/[Android Open Source Project,role=external,window=_blank]
* providing fine-grained (per-branch, per-repository, inheriting)
access controls, which allow a Gerrit admin to delegate permissions
to different team(-lead)s.
* facilitate code review: Gerrit offers a web view of pending code
changes, that allows for easy reading and commenting by humans. The
web view can offer data coming out of automated QA processes (eg.
CI). The permission system also includes fine grained control of who
can approve pending changes for submission to further facilitate
delegation of code ownership.
== Overview
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.
command line program, or any GUI which can invoke `git push` on behalf
of the user. Authentication and data transfer are handled through SSH
and HTTPS. Uploads are protected by the authentication,
confidentiality and integrity offered by the transport (SSH, HTTPS).
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 NoteDb.
Each Git commit created on the client desktop system is converted into
a unique change record which can be reviewed independently.
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.
`git push` command line, but typically reviewers are added in the web
interface.
Reviewers use the web interface to read the side-by-side or unified
diff of a change, and insert draft inline/file comments where
@@ -103,20 +65,16 @@ 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.
Reviewers can score the change ("vote"), 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 equivalent 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.
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
user pressing the submit button does not need to be the author of the
change.
== Infrastructure
@@ -125,18 +83,30 @@ End-user web browsers make HTTP requests directly to Gerrit's
HTTP server. As nearly all of the user interface is implemented
through PolyGerrit, the majority of these requests are transmitting
compressed JSON payloads, with all HTML being generated within the
browser. Most responses are under 1 KB.
browser.
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.
Gerrit's HTTP server side component is implemented as a standard Java
servlet, and thus runs within any link:install-j2ee.html[J2EE servlet
container]. The standard install will run inside Jetty, which is
included in the binary.
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.
End-user uploads are performed over SSH or HTTP, 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.
User authentication is handled by identity realms. Gerrit supports the
following types of authentication:
* OpenId (see link:http://openid.net/developers/specs/[OpenID Specifications,role=external,window=_blank])
* OAuth2
* LDAP
* Google accounts (on googlesource.com)
* SAML
* Kerberos
* 3rd party SSO
=== NoteDb
Server side data storage for Gerrit is broken down into two different
categories:
@@ -156,28 +126,119 @@ 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 mostly housed in the database (*1),
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 the metadata database on the same server,
to reduce administration overheads.
The Gerrit metadata contains a summary of the available changes, all
comments (published and drafts), and individual user account
information.
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 metadata is also stored in Git, with the commits marking the
historical state of metadata. Data is stored in the trees associated
with the commits, typically using Git config file or JSON as the base
format. For metadata, there are 3 types of data: changes, accounts and
groups.
* link:http://www.kernel.org/pub/software/scm/git/docs/gitrepository-layout.html[Git Repository Format,role=external,window=_blank]
* link:http://openid.net/developers/specs/[OpenID Specifications,role=external,window=_blank]
Accounts are stored in a special Git repository `All-Users`.
*1 Although an effort is underway to eliminate the use of the
database altogether, and to store all the metadata directly in
the git repositories themselves. So far, as of Gerrit 2.2.1, of
all Gerrit's metadata, only the project configuration metadata
has been migrated out of the database and into the git
repositories for each project.
Accounts can be grouped in groups. Gerrit has a built-in group system,
but can also interface to external group system (eg. Google groups,
LDAP). The built-in groups are stored in `All-Users`.
Draft comments are stored in `All-Users` too.
Permissions are stored in Git, in a branch `refs/meta/config` for the
repository. Repository configuration (including permissions) supports
single inheritance, with the `All-Projects` repository containing
site-wide defaults.
Code review metadata is stored in Git, alongside the code under
review. Metadata includes change status, votes, comments. This review
metadata is stored in NoteDb along with the submitted code and code
under review. Hence, the review history can be exported with `git
clone --mirror` by anyone with sufficient permissions.
== Permissions
Permissions are specified on branch names, and given to groups. For
example,
```
[access "refs/heads/stable/*"]
push = group Release-Engineers
```
this provides a rule, granting Release-Engineers push permission for
stable branches.
There are fundamentally two types of permissions:
* Write permissions (who can vote, push, submit etc.)
* Read permissions (who can see data)
Read permissions need special treatment across Gerrit, because Gerrit
should only surface data (including repository existence) if a user
has read permission. This means that
* The git wire protocol support must omit references from
advertisement if the user lacks read permissions
* Uploads through the git wire protocol must refuse commits that are
based on SHA1s for data that the user can't see.
* Tags are only visible if their commits are visible to user through a
non-tag reference.
Metadata (eg. OAuth credentials) is also stored in Git. Existing
endpoints must refuse creating branches or changes that expose these
metadata or allow changes to them.
=== Indexing
Almost all data is stored as Git, but Git only supports fast lookup by
SHA1 or by ref (branch) name. Therefore Gerrit also has an indexing
system (powered by Lucene by default) for other types of queries.
There are 4 indices:
* Project index - find repositories by name, parent project, etc.
* Account index - find accounts by name, email, etc.
* Group index - find groups by name, owner, description etc.
* Change index - find changes by file, status, modification date etc.
The base entities are characterized by SHA1s. Storing the
characterizing SHA1s allows detection of stale index entries.
== Plug-in architecture
Gerrit has a plug-in architecture. Plugins can be installed by
dropping them into $site_directory/plugins, or at runtime through
plugin SSH commands, or the plugin REST API.
=== Backend plugins
At runtime, code can be loaded from a `.jar` file. This code can hook
into predefined extension points. A common use of plugins is to have
Gerrit interoperate with site-specific tools, such as CI-systems or
issue trackers.
// list some notable extension points, and notable plugins
// link to plugin development
Some backend plugins expose the JVM for scripting use (eg. Groovy,
Scala), so plugins can be written without having to setup a Java
development environment.
// Luca to expand: how do script plugins load their scripts?
=== Frontend plugins
The UI can be extended using Frontend plugins. This is useful for
changing the look & feel of Gerrit, but it can also be used to surface
data from systems that aren't integrated with the Gerrit backend, eg.
CI systems or code coverage providers.
// FE team to write a bit more:
// * how to load ?
// * XSRF, CORS ?
== Internationalization and Localization
@@ -189,14 +250,11 @@ 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.
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.
== Accessibility Considerations
// UI team to rewrite this.
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.
@@ -215,7 +273,9 @@ provide hints to screen readers.
== Browser Compatibility
Supporting non-JavaScript enabled browsers is a non-goal for Gerrit.
Gerrit requires a JavaScript enabled browser.
// UI team to add section on minimum browser requirements.
As Gerrit is a pure JavaScript application on the client side, with
no server side rendering fallbacks, the browser must support modern
@@ -223,54 +283,19 @@ 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.
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.
All of the content stored within Gerrit is also available through
other means, such as gitweb or the `git://` protocol. Any existing
search engine crawlers can index the server-side HTML served by a code
browser, and thus can index the majority of the changes which might
appear in Gerrit. Therefore the lack of support for most search engine
crawlers is a non-issue for most Gerrit deployments.
== Product Integration
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 optionally surfaces links to HTML pages in a code browser. The
links are configurable, and Gerrit comes with a built-in code browser,
called Gitiles.
Gerrit integrates with some types of corporate single-sign-on (SSO)
solutions, typically by having the SSO authentication be performed
@@ -290,16 +315,17 @@ they choose.
Gerrit does not integrate with any Google service, or any other
services other than those listed above.
Plugins (see above) can be used to drive product integrations from the
Gerrit side. Products that support Gerrit explicitly can use the REST
API or the SSH API to contact Gerrit.
== Privacy Considerations
Gerrit stores the following information per user account:
* Full Name
* Preferred Email Address
* Mailing Address '(Optional, Encrypted)'
* Country '(Optional, Encrypted)'
* Phone Number '(Optional, Encrypted)'
* Fax Number '(Optional, Encrypted)'
The full name and preferred email address fields are shown to any
site visitor viewing a page containing a change uploaded by the
@@ -325,271 +351,145 @@ project's mailing list archives.
The user's name and email address is stored unencrypted in the
link:config-accounts.html#all-users[All-Users] repository.
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.
== Spam and Abuse Considerations
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.
There is no spam protection for the Git protocol upload path.
Uploading a change successfully requires a pre-existing account, and a
lot of up-front effort.
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.
Gerrit makes no attempt to detect spam changes or comments in the web
UI. To post and publish a comment a client must sign in and then use
the XSRF protected JSON-RPC interface to publish the draft on an
existing change record.
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 fewer 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.
== Latency
Gerrit targets for sub-250 ms per page request, mostly by using
very compact JSON payloads between client and server. However, as
most of the serving stack (network, hardware, metadata
database) is out of control of the Gerrit developers, no real
guarantees can be made about latency.
Absence of SPAM handling is based upon the idea that Gerrit caters to
a niche audience, and will therefore be unattractive to spammers. In
addition, it is not a factor for corporate, on-premise deployments.
== Scalability
Gerrit is designed for a very large scale open source project, or
large commercial development project. Roughly this amounts to
parameters such as the following:
Gerrit supports the Git wire protocol, and an API (one API for HTTP,
and one for SSH).
.Design Parameters
[options="header"]
|======================================================
|Parameter | Default Maximum | Estimated Maximum
|Projects | 1,000 | 10,000
|Contributors | 1,000 | 50,000
|Changes/Day | 100 | 2,000
|Revisions/Change | 20 | 20
|Files/Change | 50 | 16,000
|Comments/File | 100 | 100
|Reviewers/Change | 8 | 8
|======================================================
The git wire protocol does a client/server negotiation to avoid
sending too much data. This negotation occupies a CPU, so the number
of concurrent push/fetch operations should be capped by the number of
CPUs.
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
available to the JVM and the relevant cache.*.memoryLimit variables
are increased from their defaults.
=== Discussion
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
link:config-gerrit.html#cache.name.memoryLimit[`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,
link:config-gerrit.html#cache.name.memoryLimit[`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
link:config-gerrit.html#cache.name.memoryLimit[`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.
=== CPU Usage - Web UI
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/file 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).
----
QPD = Changes_Day * Revisions_Change * Reviewers_Change * (4 + F + F * C)
= 2,000 * 2 * 1 * (4 + 10 + 10 * 4)
= 216,000
QPS = QPD / 8_Hours / 60_Minutes / 60_Seconds
= 7.5
----
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.
=== CPU Usage - Git over SSH/HTTP
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, link:pgm-daemon.html[replica 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
Clients on slow network connections 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.
Deployments for large, distributed companies can replicate Git data to
read-only replicas to offload fetch traffic. The read-only replicas
should also serve this data using Gerrit to ensure that permissions
are obeyed.
=== Disk Usage
The API serves requests of varying costs. Requests that originate in
the UI can block productivity, so care has been taken to optimize
these for latency, using the following techniques:
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 maximum 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.
* Async calls: the UI becomes responsive before some UI elements
finished loading
* Caching: metadata is stored in Git, which is relatively expensive to
access. This is sped up by multiple caches. Metadata entities are
stored in Git, and can therefore be seen as immutable values keyed
by SHA1, which is very amenable to caching. All SHA1 keyed caches
can be persisted on local disk.
The size (memory, disk) of these caches should be adapted to the
instance size (number of users, size and quantity of repositories)
for optimal performance.
Git does not impose fundamental limits (eg. number of files per
change) on data. To ensure stability, Gerrit configures a number of
default limits for these.
// add a link to the default settings.
=== Scaling team size
A team of size N has N^2 possible interactions. As a result, features
that expose interactions with activities of other team members has a
quadratic cost in aggregate. The following features scale poorly with
large team sizes:
* the change screen shows conflicting changes by default. This data is
cached, but updates to pending changes cause cache misses. For a
single change, the amount of work is proportional to the number of
pending changes, so in aggregate, the cost of this feature is
quadratic in the team size.
* the change screen shows if a change is mergeable to the target
branch. If the target branch moves quickly (large developer team),
this causes cache misses. In aggregate, the cost of this feature is
also quadratic.
Both features should be turned off for repositories that involve 1000s
of developers.
=== Browser performance
// say something about browser performance tuning.
=== Real life numbers
Gerrit is designed for very large projects, both open source and
proprietary commercial projects. For a single Gerrit process, the
following limits are known to work:
.Observed maximums
[options="header"]
|======================================================
|Parameter | Maximum | Deployment
|Projects | 50,000 | gerrithub.io
|Contributors | 150,000 | eclipse.org
|Bytes/repo | 100G | Qualcomm internal
|Changes/repo | 300k | Qualcomm internal
|Revisions/Change | 300 | Qualcomm internal
|Reviewers/Change | 87 | Qualcomm internal
|======================================================
// find some numbers for these stats:
// |Files/repo | ? |
// |Files/Change | ? |
// |Comments/Change | ? |
// |max QPS/CPU | ? |
Google runs a horizontally scaled deployment. We have seen the
following per-JVM maximums:
.Observed maximums (googlesource.com)
[options="header"]
|======================================================
|Parameter | Maximum | Deployment
|Files/repo | 500,000 | chromium-review
|Bytes/repo | 12G | chromium-review
|Changes/repo | 500k | chromium-review
|Revisions/Change | 1900 | chromium-review
|Files/Change | 10,000| android-review
|Comments/Change | 1,200 | chromium-review
|======================================================
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).
== Redundancy & Reliability
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 is structured as a single JVM process, reading and writing to a
single file system. If there are hardware failures in the machine
running the JVM, or the storage holding the repositories, there is no
recourse; on failure, errors will be returned to the client.
Gerrit largely assumes that the metadata 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.
Deployments needing more stringent uptime guarantees can use
replication/multi-master setup, which ensures availability and
geographical distribution, at the cost of slower write actions.
Due to the relatively small scale described above, it is very likely
that the Git filesystem and 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.
// TODO: link.
=== Backups
@@ -603,7 +503,8 @@ Amazon S3 blob storage service.
== Logging Plan
Gerrit does not maintain logs on its own.
Gerrit stores Apache style HTTPD logs, as well as ERROR/INFO messages
from the Java logger, under `$site_dir/logs/`.
Published comments contain a publication date, so users can judge
when the comment was posted and decide if it was "recent" or not.