Gate Failure Triage

This page provides guidelines for spotting and assessing neutron gate failures. Some hints for triaging failures are also provided.

Spotting Gate Failures

This can be achieved using several tools:

For checking gate failures with opensearch please see documentation. The following query will return failures for a specific job:

> build_status:FAILURE AND message:Finished AND build_name:”check-tempest-dsvm-neutron” AND build_queue:”gate”

And divided by the total number of jobs executed:

> message:Finished AND build_name:”check-tempest-dsvm-neutron” AND build_queue:”gate”

It will return the failure rate in the selected period for a given job. It is important to remark that failures in the check queue might be misleading as the problem causing the failure is most of the time in the patch being checked. Therefore it is always advisable to work on failures occurred in the gate queue. However, these failures are a precious resource for assessing frequency and determining root cause of failures which manifest in the gate queue.

The step above will provide a quick outlook of where things stand. When the failure rate raises above 10% for a job in 24 hours, it’s time to be on alert. 25% is amber alert. 33% is red alert. Anything above 50% means that probably somebody from the infra team has already a contract out on you. Whether you are relaxed, in alert mode, or freaking out because you see a red dot on your chest, it is always a good idea to check on daily bases the elastic-recheck pages.

Under the gate pipeline tab, you can see gate failure rates for already known bugs. The bugs in this page are ordered by decreasing failure rates (for the past 24 hours). If one of the bugs affecting Neutron is among those on top of that list, you should check that the corresponding bug is already assigned and somebody is working on it. If not, and there is not a good reason for that, it should be ensured somebody gets a crack at it as soon as possible. The other part of the story is to check for uncategorized failures. This is where failures for new (unknown) gate breaking bugs end up; on the other hand also infra error causing job failures end up here. It should be duty of the diligent Neutron developer to ensure the classification rate for neutron jobs is as close as possible to 100%. To this aim, the diligent Neutron developer should adopt the procedure outlined in the following sections.

Troubleshooting Tempest jobs

  1. Open logs for failed jobs and look for logs/testr_results.html.gz.

  2. If that file is missing, check console.html and see where the job failed.
    1. If there is a failure in devstack-gate-cleanup-host.txt it’s likely to be an infra issue.

    2. If the failure is in devstacklog.txt it could a devstack, neutron, or infra issue.

  3. However, most of the time the failure is in one of the tempest tests. Take note of the error message and go to opensearch.

  4. On opensearch, search for occurrences of this error message, and try to identify the root cause for the failure (see below).

  5. File a bug for this failure, and push an Elastic Recheck Query for it.

  6. If you are confident with the area of this bug, and you have time, assign it to yourself; otherwise look for an

    assignee or talk to the Neutron’s bug czar to find an assignee.

Troubleshooting functional/fullstack job

  1. Go to the job link provided by Zuul CI.

  2. Look at logs/testr_results.html.gz for which particular test failed.

  3. More logs from a particular test are stored at logs/dsvm-functional-logs/<path_of_the_test> (or dsvm-fullstack-logs for fullstack job).

  4. Find the error in the logs and search for similar errors in existing launchpad bugs. If no bugs were reported, create a new bug report. Don’t forget to put a snippet of the trace into the new launchpad bug. If the log file for a particular job doesn’t contain any trace, pick the one from testr_results.html.gz.

  5. Create an Elastic Recheck Query

Troubleshooting Grenade jobs

Grenade is used in the Neutron gate to test every patch proposed to Neutron to ensure it will not break the upgrade process. Upgrading from the N-1 to the N branch is constantly being tested. So if you send patch to the Neutron master branch Grenade jobs will first deploy Neutron from the last stable release and then upgrade it to the master branch with your patch. Details about how Grenade works are available in the documentation.

In Neutron CI jobs that use Grenade are run in the multinode jobs configuration which means that we have deployed OpenStack on 2 VMs:

  • one called controller which is in an “all in one” node so it runs neutron-server, as well as the neutron-ovs-agent and nova-compute services,

  • one called compute1 which runs only services like nova-compute and neutron-ovs-agent.

Neutron supports that neutron-server in N version will always work with the agents which runs in N-1 version. To test such scenario all our Grenade jobs upgrade OpenStack services only on the controller node. Services which run on the compute1 node are always run with the “old” release during that job.

Debugging of failures in the Grenade job is very similar to debugging any other Tempest based job. The difference is that in the logs of the Grenade job, there is always “logs/old” and “logs/new” directories which contain Devstack logs from each run of the Devstack’s stack.sh script. In the “logs/grenade.sh_log.txt” file there is a full log of the grenade.sh run and you should always start checking failures from that file. Logs of the Neutron services for “old” and “new” versions are in the same files, like, for example, “logs/screen-q-svc.txt” for neutron-server logs. You will find in that log when the service was restarted - that is the moment when it was upgraded by Grenade and it is now running the new version.

Advanced Troubleshooting of Gate Jobs

As a first step of troubleshooting a failing gate job, you should always check the logs of the job as described above. Unfortunately, sometimes when a tempest/functional/fullstack job is failing, it might be hard to reproduce it in a local environment, and might also be hard to understand the reason of such a failure from only reading the logs of the failed job. In such cases there are some additional ways to debug the job directly on the test node in a live setting.

This can be done in two ways:

  1. Using the remote_pdb python module and telnet to directly access the python debugger while in the failed test.

    To achieve this, you need to send a Do not merge patch to gerrit with changes as described below:

    • Add an iptables rule to accept incoming telnet connections to remote_pdb. This can be done in one of the ansible roles used in the test job. Like for example in neutron/roles/configure_functional_tests file for functional tests:

      sudo iptables -I openstack-INPUT -p tcp -m state --state NEW -m tcp --dport 44444 -j ACCEPT
      
    • Increase the OS_TEST_TIMEOUT value to make the test wait longer when remote_pdb is active to make debugging easier. This change can also be done in the ansible role mentioned above:

      export OS_TEST_TIMEOUT=999999
      

      Please note that the overall job will be limited by the job timeout, and that cannot be changed from within the job.

    • To make it easier to find the IP address of the test node, you should add to the ansible role so it prints the IPs configured on the test node. For example:

      hostname -I
      
    • Add the package remote_pdb to the test-requirements.txt file. That way it will be automatically installed in the venv of the test before it is run:

      $ tail -1 test-requirements.txt
      remote_pdb
      
    • Finally, you need to import and call the remote_pdb module in the part of your test code where you want to start the debugger:

      $ diff --git a/neutron/tests/fullstack/test_connectivity.py b/neutron/tests/fullstack/test_connectivity.py
      index c8650b0..260207b 100644
      --- a/neutron/tests/fullstack/test_connectivity.py
      +++ b/neutron/tests/fullstack/test_connectivity.py
      @@ -189,6 +189,8 @@ class
      TestLinuxBridgeConnectivitySameNetwork(BaseConnectivitySameNetworkTest):
              ]
      
           def test_connectivity(self):
      +        import remote_pdb; remote_pdb.set_trace('0.0.0.0', port=44444)
      +
      self._test_connectivity()
      

      Please note that discovery of public IP addresses is necessary because by default remote_pdb will only bind to the 127.0.0.1 IP address. Above is just an example of one of possible method, there could be other ways to do this as well.

    When all the above changes are done, you must commit them and go to the Zuul status page to find the status of the tests for your Do not merge patch. Open the console log for your job and wait there until remote_pdb is started. You then need to find the IP address of the test node in the console log. This is necessary to connect via telnet and start debugging. It will be something like:

    RemotePdb session open at 172.99.68.50:44444, waiting for connection ...
    

    An example of such a Do not merge patch described above can be found at https://review.opendev.org/#/c/558259/.

    Please note that after adding new packages to the requirements.txt file, the requirements-check job for your test patch will fail, but it is not important for debugging.

  2. If root access to the test node is necessary, for example, to check if VMs have really been spawned, or if router/dhcp namespaces have been configured properly, etc., you can ask a member of the infra-team to hold the job for troubleshooting. You can ask someone to help with that on the openstack-infra IRC channel. In that case, the infra-team will need to add your SSH key to the test node, and configure things so that if the job fails, the node will not be destroyed. You will then be able to SSH to it and debug things further. Please remember to tell the infra-team when you finish debugging so they can unlock and destroy the node being held.

The above two solutions can be used together. For example, you should be able to connect to the test node with both methods:

  • using remote_pdb to connect via telnet;

  • using SSH to connect as a root to the test node.

You can then ask the infra-team to add your key to the specific node on which you have already started your remote_pdb session.

Root Causing a Gate Failure

Time-based identification, i.e. find the naughty patch by log scavenging.

Filing An Elastic Recheck Query

The elastic recheck page has all the current open ER queries. To file one, please see the ER Wiki.