Almost everything that exists on a Fuchsia system is a Fuchsia package. Whether it is immediately apparent or not almost everything you see on Fuchsia lives in a package. This document will cover the basics of a package-driven workflow where you build a package and push it to a Fuchsia device which is reachable via IP from your development host.
The host and target must be able to communicate over IP. In particular it must be possible to SSH from the development host to the target device, and the target device must be able to connect via TCP to the development host on port 8083. The SSH connection is used to issue commands to the target device.
The development host will run a simple, static file, HTTP server which makes the updates available to the target. This HTTP server is part of the Fuchsia source code and built automatically.
The target is instructed to look for changes on the development host via a couple of commands that are run manually. When the update system on the target sees these changes it will fetch the new software from the HTTP server running on the host. The new software will be available until the target is rebooted.
To build a package containing the required code, a package type build rule is used. If one of these needs to be created for the target package, consult the reference page for this. Some build rule types are actually extensions of the package rule type, for example
flutter_app extends the package type.
Once an appropriate build rule is available the target package can be re-generated by running
The Fuchsia source contains a simple HTTP server which serves static files. The build generates a TUF file tree which is served.
The update agent on the target does not initially know where to look for updates. To connect the agent on the target to the HTTP server running on the development host, it must be told the IP address of the development host. The host HTTP server is started and the update agent is configured by calling
fx serve -v or
fx serve-updates -v.
fx serve will run both the bootserver and the update server and is often what people use.
fx serve-updates runs just the update server. In both cases,
-v is recommended because the command will print more output which may assist with debugging. If the host connects successfully to the target you will see the message
Ready to push packages! in the shell on your host.
The update agent on the target will remain configured until it is repaved or persistent data is lost. The host will attempt to reconfigure the update agent when the target is rebooted.
Packages in Fuchsia are not “installed”, they are cached on an as needed basis. There are two collections of packages on a Fuchsia system:
base The base package set is a group of software critical to proper system function that must remain congruent. This set of software can only be updated by performing a whole system update, typically referred to as OTA, described below. This is updated using
ephemeral software The ephemeral software is provided to the system in one of two ways, either as a member of the “cache” build group, or entirely ephemerally. Typically users configure software in the Fuchsia build as ephemerally available by adding that package to the
fx set line such as
--with //examples/rolldice. Ephemeral software, whether in “cache” or entirely ephemeral, is always updated whenever a component is launched with a package URL. For example, if a user executes
fx run fuchsia-pkg://fuchsia.com/rolldice#meta/rolldice.cmx, the rolldice latest rolldice package will be cached before execution.
Sometimes there may be many packages changed or the kernel may change or there may be changes in the system package. To get kernel changes or changes in the system package an OTA or pave is required as base packages are immutable for the runtime of a system. An OTA update will usually be faster than paving or flashing the device.
fx ota asks the target device to perform an update from any of the update sources available to it. To OTA update a build made on the dev host to a target on the same LAN, first build the system you want. If
fx serve [-v] isn't already running, start it so the target can use the development host as an update source. The
-v option will show more information about the files the target is requesting from the host. If the
-v flag was used there should be a flurry of output as the target retrieves all the new files. Following completion of the OTA the device will reboot.
fx serve -v(to run the update server for both build-push and ota)
fx serve-updates -v(to run only the update server, not the bootserver)
fx run <component-url>(each time a change is made you want to run)
fx test <component-url>(to build and run tests)
fx ota(to trigger a full system update and reboot)
Every update pushed is stored in the content-addressed file system, blobfs. Following a reboot the updated packages may not be available because the index that locates them in blobfs is only held in RAM. The system currently does not garbage collect inaccessible or no-longer-used packages (having garbage to collect is a recent innovation!), but will eventually.
fx gc will reboot the target device and then evict all old ephemeral software from the device, freeing up space.
If the package being updated hosts a service managed by Fuchsia that service may need to be restarted. Rebooting is undesirable both because it is slow and because the package will revert to the version paved on the device. Typically a user can terminate one or more running components on the system, either by asking the component to terminate gracefully, or by forceufully stopping the component using
fx shell killall <component-name>. Upon reconnection to the component services, or by invocation via
fx run or
fx test, new versions available in the package server will be cached before launch.
Packaging and pushing code that lives outside the Fuchsia tree is possible, but will require more work. The Fuchsia package format is quite simple. It consists of a metadata file describing the package contents which is described in more detail in the Fuchsia package documentation. The metadata file is added to a TUF file tree and each of the contents are named after their Merkle root hash and put in a directory at the root of the TUF file tree called ‘blobs’.