Comprehensive, High-Level Design Description
MozPool is a tool for managing a pool of untrustworthy mobile devices. It is deployed as a single system, but comprised of several distinct components for design simplicity.
It shouldn't cause too much confusion that the top-level component is also known as MozPool. It's just such a great name.
MozPool is responsible for matching requests with devices. A new request is submitted by a client with parameters for acceptable devices (may be as broad as "anything" or as narrow as "this panda" and the expected condition of that device (Android suitable for Fennec, or a particular B2G image, or booted to the live image for diagnostic purposes). Clients can be automated test systems (Buildbot, Autophone) or flesh-and-blood users.
Requests are filled by matching them with a single device. Once that match is made and returned to the client, the request stays around as a form of reservation. Reservations time out if they are not renewed periodically, where the period is specified by the client (so flesh-and-blood users can reserve a device for a day or two, while automated systems can use 30 minutes or something smaller).
When matching a request to a device, MozPool picks a device itself, but relies on LifeGuard to keep information about the available devices up to date, and to put the requested device in the desired state. If LifeGuard fails to set up the device as desired, MozPool is responsible for picking another device that satisfies the request, or indicating failure to the client, if the parameters of the request cannot be satisfied.
MozPool also provides various statistics and reports as needed to maintain the health of the pool. These include summaries of the status of devices by type (where status is divided into simple categories like "in use", "idle", "processing", and "failed"); and lists of devices in known failure states requiring human remediation.
In the initial design, MozPool is entirely reactive, but the design does not preclude predictive or proactive operations, e.g., balancing the distribution of images on spare devices, predictively installing B2G images, etc.
LifeGuard deals only with devices. It actively tracks the state of every device, and handles requests from MozPool to change the state of a device. These requests are conditional: "please change to state Y iff the device is currently in state X"; the state-change request fails if the condition is not met.
Most states for a device involve periodic checks from LifeGuard. Idle devices are checked periodically, and failure conditions rectified.
BMM, short for Black Mobile Magic, is the lowest-level component, and handles technical operations on devices as requested from LifeGuard. The available operations are power-cycling a device; PXE-booting a device; and running commands on a device via SUTAgent. BMM includes TFTP and HTTP services to allow a device to be booted into a Linux live-boot environment, and scripts run there to perform whatever actions are appropriate.
Specific scripts will implement actions required by LifeGuard: install Android, install a B2G image, run an SSH server in maintenance mode, run system checks, etc.
BMM abstracts away the details of how power is controlled for each device, as well as the particulars of boot images for specific hardware.
As much logging as possible is funneled through syslog and into the mysql database, to help with debugging.
Logs are expired after some time by the database itself (see
The Mozilla inventory (https://inventory.mozilla.org) is the source of truth from which the list of devices is derived. The database is automatically synchronized with inventory periodically.
Each device is assigned, in inventory, to a specific mobile-imaging server. In general, that server is "close" to the device, physically or virtually.
All three major components are implemented in the same Python daemon, running web services based on web.py. An instance of this daemon runs on each mobile-imaging server.
The daemon runs background processes in separate threads. In particular, various operations poll for status.
There is no front-end load balancer. If an imaging server is down or unavailable, the devices assigned to it are also unavailable, but other devices continue to be accessible.
Clients access MozPool using an HTTP API. The endpoint for that API is any mobile-imaging server, since all are configured identically. Clients should be pre-configured with a list of servers, and retry servers in random order until successful.
The entire lifetime of each request is handled by MozPool as a formal state machine. The state is stored in the database.
All state transitions and actions are handled on the server where the request was originally made. Timeouts are handled by polling the database for requests with timeout timestamps in the past (using threads within the daemon).
If an imaging server is lost, the requests it manages become invalid when their refresh interval expires.
Boards are claimed by inserting into a correspondance table in the database, with constraints such that only one request can claim a device.
Like requests, devices are managed by LifeGuard as a formal state machine. MozPool has read-only visibility to device states for purposes of selecting devices for requests, but uses conditional requests to LifeGuard to cause state transitions (the intent being that MozPool will observe that a device is in the idle state, claim it, then ask that LifeGuard transition it from idle to rebooting; if the device has failed in the interim, LifeGuard will refuse to do so).
All state transitions and actions are handled on the server to which the device is assigned.
MozPool communicates with LifeGuard using an HTTP API, selecting the endpoint based on the assigned imaging server in the database. This may result in a MozPool server contacting itself via HTTP.
LifeGuard communicates with BMM using regular old Python function invocations.
Configuration should be based on the mozpool/config.ini.dist template. The
config can either be put in the
mozpool/config.ini, or anywhere else with
$BMM_CONFIG giving the full path.
To run the server daemon:
optionally, add a port on the command line for the HTTP server:
To install the DB schema (using the configured database):
And to install test adta
mozpool-db run testdata.py
To control relays:
relay powercycle <hostname> <bank> <relay> relay status <hostname> <bank> <relay> relay turnon <hostname> <bank> <relay> relay turnoff <hostname> <bank> <relay>
Note: do not manually adjust relays that are also under MozPool's active control!
To synchronize the internal DB with inventory:
--verbose to see what it's up to - note that it's not too fast!)
To run the tests:
- install mock
- install paste