V2 provides detailed control over timeouts within testjobs. The job has an over-arching timeout, each action has a default runtime timeout, each connection has a default timeout to get a reply and individual actions within the pipeline can have the action or connection timeout modified in the test definition or in the device configuration.
Automation and Timeouts¶
Why must timeouts exist?¶
Automation needs to be able to cope with failures during operation and one of the most common failure modes is that a command never returns because some part of the setup for that command has not been done, failed during operation or is not supported for some reason.
It is not usually the operation which times out which is the reason for the failure. Typically, a timeout results from one or more previous operations failing. A test job which fails due to a timeout must always be investigated - simply extending the timeout is the wrong approach. Equally, long timeouts are directly harmful to other users, choose your timeouts carefully and regularly evaluate whether existing timeouts can be shortened.
When operating the same test interactively, a human will typically notice that a step has failed and either not attempt the operation that would later time out or recognize that the operation should have completed by some point in time and intervene. It is not always possible to check the success or failure of operations within a test in an automated fashion - command outputs change from one distribution to another or from one version of a package to another.
The timeout is a necessary part of automation, it allows test jobs to fail instead of holding on to the job and the device indefinitely.
How long should an operation wait?¶
The best guidance here is to assess how the operation works when you use the device interactively. Compare over a few runs of the operation and then use a timeout which is slightly longer than the longest successful operation by rounding up to the nearest whole minute or hour.
Timeouts do not need to be precise, 2 minutes is better than 90 seconds, but yet must not be excessive. If an operation has routinely taken 5 minutes to succeed previously and has now suddenly taken 30 minutes, this needs to be investigated. It could easily be a kernel bug, hardware fault, infrastructure fault or test job error. Increase timeouts gradually and keep a sense of perspective of just what is reasonable to expect an operation to require.
Do not simply transfer timeouts from V1 jobs! V1 timeouts do not have the same structure and cannot be easily mapped to individual actions or operations within an action.
Operations involving third party services¶
When downloading, uploading or transferring data using third party services it can be hard to estimate a reasonable timeout. If test jobs start to fail during such operations, investigate whether the connection to the third party service can be improved, cached or fixed.
Duration of actions¶
The duration of every action in a test job is tracked and recorded. This allows test writers to look at other similar jobs and evaluate the actual duration of any operation within that testjob. Equally, it allows lab admins to compare your timeouts against the actual duration of the operation. If your jobs start to fail and sit idle for long periods waiting for a timeout, you have the information to hand to fix the timeouts yourself before you get a prompt from the admins.
Test shell timeouts¶
The timeout used by the test action is a single value covering all test operations. Actual durations are still tracked and recorded, so excessive timeouts still need to be addressed.
For the test writer, the timeout is expressed as a single integer value of:
There is no need to specify sub-divisions or to overflow. Instead of
minutes: 2 and instead of trying to specify two and a half
minutes, just use
minutes: 3. Using
hours: 2 when only
is required is likely to get you a warning from the admins but using
10 instead of
seconds: 600 is strongly recommended.
Although timeouts support
days, you need to have a very good reason to
set such a timeout to avoid being accused of denying access to the device to
other users (including the special
lava-health user which is used to submit
The entire test job has a single over-arching timeout. This means that no matter how long any action or connection timeout is set within the test job, if the test job duration increases above the job timeout then the slave will terminate the job and set the status as Incomplete.
The first reason for this timeout is so that individual actions or connections can have freedom to set timeouts but the testjob still fails if more than one or two of the operations take significantly longer than anticipated.
The second reason for a job timeout is that it allows the UI to derive an estimate of how long the job will take to inform other users who may be waiting for their jobs to start on the busy devices.
timeouts: job: minutes: 15 action: minutes: 5 connection: minutes: 4 actions: u-boot-commands: minutes: 3 connections: lava-test-shell: minutes: 4
timeouts block specifies the job timeout, as well as the
Default action timeouts (5 minutes in this example) and
Default connection timeouts (4 minutes in this example).
Summary of the example job timeouts¶
The test job will not take longer than 15 minutes or it will timeout. This will happen irrespective of which action is currently running or how much time that action has before it would timeout.
No one action (deploy, boot or test) will take longer than 5 minutes or that action will timeout. Each operation within the action (the action class) will pass on the remaining time to the next operation. Enable the debug logs at the top of the log page to see this as a decreasing
timeoutvalue with each
start: 1.3.4 compress-overlay (timeout 00:04:06) end: 1.3.4 compress-overlay (duration 00:00:03) start: 1.3.5 persistent-nfs-overlay (timeout 00:04:03)
No one connection will take longer than 2 minutes or the action will timeout. Connection timeouts are between prompts, so this is the maximum amount of time that any operation within the action can take before the action determines that there is not going to be any more output and to fail as a timeout. Actions typically include multiple connections, each with the same timeout. Connection timeouts are not affected by previous connections, each time a command is sent, the action expects to find the prompt again within the same connection timeout.
All timeouts in this top level section can be overridden later in the test job definition.
Default action timeouts¶
An action timeout covers the entire operation of all operations performed by that action. Check the V2 logs for lines like:
start: 1.1.1 http_download (timeout 00:05:00)
end: 1.1.1 file_download (duration 00:00:25)
The action timeout
00:05:00 comes from this part of the job definition:
timeouts: job: minutes: 15 action: minutes: 5
The complete list of actions for any test job is available from the job definition page, on the pipeline tab.
Not all actions in any one pipeline will perform any operations.
Action classes are idempotent and can skip operations depending on the
parameters of the testjob. Hence some actions will show a duration of
Default connection timeouts¶
A connection timeout covers each single operation of sending a command to a device and getting a response back from that device. A new connection timeout is used for each operation of sending a command to the device. For example, when sending a list of commands to a bootloader, each complete line has the same connection timeout which is reset back to zero for the subsequent line.
Connection timeouts can be much shorter than action timeouts, especially if the action needs to send multiple lines of commands.
Inheriting timeouts from the device configuration¶
In addition, individual device types can set an action override or connection override for all pipelines using devices of that type. This is to allow for certain devices which need to initialize certain hardware that takes longer than most other devices with similar support.
Details of these timeouts can be seen on the device type page on the Support tab and can be overridden using the overrides in the test job.
The actual timeout for each action is computed by taking the device configuration and overriding the values with the timeouts from the job definition. The timeout will be the first defined value in: Action block overrides, Individual action overrides and Default action timeouts.
Individual action overrides¶
For fine-grained control over action timeouts, individual actions can be named in the timeout block at the top of the test job submission and assigned a specific timeout which can be longer or shorter than the default or the action block override.
timeouts: actions: http-download: minutes: 2
Individual connection overrides¶
For fine-grained control over connection timeouts, individual actions can be named in the timeout block at the top of the test job submission and assigned a specific connection timeout which can be longer or shorter than the default.
timeouts: connections: http-download: minutes: 2
Action block overrides¶
The test job submission action blocks, (
also have timeouts. These will override the default action timeout for all
actions within that block. Action blocks are identified by the start of the
action level and the timeout value is set within that action block:
actions: - deploy: timeout: minutes: 3
The default timeout for each action within this block will be set to the specified value.
The timeout for individual actions in a block can also be redefined in the timeouts section within the block:
actions: - deploy: timeout: minutes: 5 timeouts: http-download: minutes: 1
Skipping a test shell timeout¶
In some cases, a test shell action is known to hang or otherwise cause a timeout.
If the device is capable of booting twice in a single test job
(deploy, boot, test, boot, test) then the first test action
timeout can set
skip: true which stops the job finishing as
Incomplete if the timeout occurs. The job will then continue to the
second boot action, allowing the device to reset to a known state and
start the second test action.
If the first test action does not timeout, the job completes that test action and executes the second boot action as normal.
Test writers should consider using utilities like the
command inside their own test shell scripts to retain control within
the currently executing test shell scripts. Skipping a test shell
timeout is intended for tests which may cause a kernel panic or other
deadlock of the currently executing test shell.
There are limitations to what can be achieved here:
Lava Test Shell is not re-entrant - it is not possible to restart or return to the previously executing test shell. The second test action is a separate test shell and the boot action must be defined by the test writer at submission.
Timeouts in MultiNode test jobs cannot be skipped.
The timeout itself must still occur - the test job must wait, after the error has occurred, until the timeout. Ensure that the timeout value is still long enough to cover the actual execution time if the test shell action did not hang or fail.
This support will not protect the DUT in the case of a destructive test shell failure. If the test shell action simply takes too long because, for example, a parameter has been missed and the script is deleting all of
/home/test/usr/lib/, the DUT will likely fail to reboot without a new deployment regardless of the skip support.
This support is intended for predictable test shell errors. Support must be planned into the test job before submission. Design your test shell definition and test result handling carefully. Any expected test case results from the test shell definition which might have occurred after the timeout will be completely missing. It is recommended to put the operation which is expected to fail as the last command in the test shell definition before the test action would normally end.
Timeouts are immediate, aggressive and external to the test shell. There is no opportunity for the active test shell to respond or handle the timeout. The currently executing process will disappear and filesystems will not be unmounted - power is simply removed from the DUT. This could affect the ability of the DUT to execute the next test shell after a reboot, for example if the filesystem cannot be mounted or the previous test action failed in the middle of an operation which relies on filesystem locks.
Test writers are wholly responsible for cleaning up any artefacts of the failed test shell at the start of the second test shell. For example:
if the first test shell fails when making persistent changes to the filesystem(s) on the DUT, filesystem corruption is possible which could cause the second boot action to fail.
Test actions which install packages are likely to leave stale lock files in place, incompletely installed packages and other breakage.
Use portable test shell definitions and ensure the integrity of the packages on the DUT before trying to make more persistent changes in the second test action block. Assume that the second test action will start in a broken system if no deploy action is specified before the second boot action.
For some devices a deploy action is also needed to get to a point where the device will boot successfully. (This is also a way of ensuring that filesystem corruption issues are avoided - by re-deploying the filesystem itself in a known clean state.)
- test: timeout: minutes: 5 skip: true definitions: # ... rest of the first test action block - boot: timeout: minutes: 2 # ... rest of the second boot action block - test: timeout: minutes: 5 definitions: # ... rest of the second test action block