![]() 12 hits in 7 days, check and gate, all failures. This is only failing on the cells v1 job. There are no clear service logs that are for this failure as it's probably some obscure race in cellsv1 state tracking. Change-Id: Ibc3d159e61f20f1b0975247d4148aa45e87821b1 Related-Bug: #1620761 |
||
---|---|---|
doc/source | ||
elastic_recheck | ||
queries | ||
tools | ||
web | ||
.coveragerc | ||
.gitignore | ||
.gitreview | ||
.testr.conf | ||
CONTRIBUTING.rst | ||
LICENSE | ||
MANIFEST.in | ||
README.rst | ||
babel.cfg | ||
bindep.txt | ||
elasticRecheck.conf.sample | ||
recheckwatchbot.yaml | ||
requirements.txt | ||
setup.cfg | ||
setup.py | ||
test-requirements.txt | ||
tox.ini | ||
web_server.py |
README.rst
elastic-recheck
"Use ElasticSearch to classify OpenStack gate failures"
- Open Source Software: Apache license
Idea
Identifying the specific bug that is causing a transient error in the gate is difficult. Just identifying which tempest test failed is not enough because a single tempest test can fail due to any number of underlying bugs. If we can find a fingerprint for a specific bug using logs, then we can use ElasticSearch to automatically detect any occurrences of the bug.
Using these fingerprints elastic-recheck can:
- Search ElasticSearch for all occurrences of a bug.
- Identify bug trends such as: when it started, is the bug fixed, is it getting worse, etc.
- Classify bug failures in real time and report back to gerrit if we find a match, so a patch author knows why the test failed.
queries/
All queries are stored in separate yaml files in a queries directory
at the top of the elastic-recheck code base. The format of these files
is ######.yaml (where ###### is the launchpad bug number), the yaml
should have a query
keyword which is the query text for
elastic search.
Guidelines for good queries:
Queries should get as close as possible to fingerprinting the root cause. A screen log query (e.g.
tags:"screen-n-net.txt"
) is typically better than a console one (tags:"console"
), as that's matching a deep failure versus a surface symptom.Queries should not return any hits for successful jobs, this is a sign the query isn't specific enough. A rule of thumb is > 10% success hits probably means this isn't good enough.
If it's impossible to build a query to target a bug, consider patching the upstream program to be explicit when it fails in a particular way.
Use the 'tags' field rather than the 'filename' field for filtering. This is primarily because of grenade jobs where the same log file shows up in the 'old' and 'new' side of the grenade job. For example,
tags:"screen-n-cpu.txt"
will query inlogs/old/screen-n-cpu.txt
andlogs/new/screen-n-cpu.txt
. Thetags:"console"
filter is also used to query inconsole.html
as well as tempest and devstack logs.Avoid the use of wildcards in queries since they can put an undue burden on the query engine. A common case where wildcards are used and shouldn't be are in querying against a specific set of
build_name
fields, e.g.gate-nova-python26
andgate-nova-python27
. Rather than usebuild_name:gate-nova-python*
, list the jobs with anOR
. For example:(build_name:"gate-nova-python26" OR build_name:"gate-nova-python27")
When adding queries you can optionally suppress the creation of
graphs and notifications by adding suppress-graph: true
or
suppress-notifcation: true
to the yaml file. These can be
used to make sure expected failures don't show up on the unclassified
page.
If the only signature available is overly broad and adding additional
logging can't reasonably make a good signature, you can also filter the
results of a query based on the test_ids that failed for the run being
checked. This can be done by adding a test_ids
keyword to
the query file and then a list of the test_ids to verify failed. The
test_id also should exclude any attrs, this is the list of attrs
appended to the test_id between '[]'. For example, 'smoke', 'slow', any
service tags, etc. This is how subunit-trace prints the test ids by
default if you're using it. If any of the listed test_ids match as
failing for the run being checked with the query it will return a match.
Since filtering leverages subunit2sql which only receives tempest test
results from the gate pipeline, this technique will only work on tempest
or grenade jobs in the gate queue. For more information about this refer
to the infra
subunit2sql documentation For example, if your query yaml file
looked like:
query: >-
message:"ExceptionA"
test_ids:
- tempest.api.compute.servers.test_servers.test_update_server_name
- tempest.api.compute.servers.test_servers_negative.test_server_set_empty_name
this will only match the bug if the logstash query had a hit for the run and either test_update_server_name or test_server_set_empty name failed during the run.
In order to support rapidly added queries, it's considered socially acceptable to approve changes that only add 1 new bug query, and to even self approve those changes by core reviewers.
Adding Bug Signatures
Most transient bugs seen in gate are not bugs in tempest associated with a specific tempest test failure, but rather some sort of issue further down the stack that can cause many tempest tests to fail.
Given a transient bug that is seen during the gate, go through the logs and try to find a log that is associated with the failure. The closer to the root cause the better.
- Note that queries can only be written against INFO level and higher log messages. This is by design to not overwhelm the search cluster.
- Since non-voting jobs are not allowed in the gate queue and e-r is
primarily used for tracking bugs in the gate queue, it doesn't spend
time tracking race failures in non-voting jobs since they are considered
unstable by definition (since they don't vote).
- There is, however, a special 'allow-nonvoting' key that can be added to a query yaml file to allow tracking non-voting job bug failures in the graph. They won't show up in the bot though (IRC or Gerrit comments).
Go to logstash.openstack.org and create an elastic search query to find the log message from step 1. To see the possible fields to search on click on an entry. Lucene query syntax is available at lucene.apache.org.
Tag your commit with a
Related-Bug
tag in the footer, or add a comment to the bug with the query you identified and a link to the logstash URL for that query search.Putting the logstash query link in the bug report is also valuable in the case of rare failures that fall outside the window of how far back log results are stored. In such cases the bug might be marked as Incomplete and the e-r query could be removed, only for the failure to re-surface later. If a link to the query is in the bug report someone can easily track when it started showing up again.
Add the query to
elastic-recheck/queries/BUGNUMBER.yaml
(All queries can be found on git.openstack.org) and push the patch up for review.
You can also help classify Unclassified failed jobs, which is an aggregation of all failed voting gate jobs that don't currently have elastic-recheck fingerprints.
Removing Bug Signatures
Old queries which are no longer hitting in logstash and are associated with fixed or incomplete bugs are routinely deleted. This is to keep the load on the elastic-search engine as low as possible when checking a job failure. If a bug marked as Incomplete does show up again, the bug should be re-opened with a link to the failure and the e-r query should be restored.
Queries that have "suppress-graph: true" in them generally should not be removed since we basically want to keep those around, they are persistent infra issues and are not going away.
Steps:
- Go to the All Pipelines page.
- Look for anything that is grayed out at the bottom which means it has not had any hits in 10 days.
- From those, look for the ones that are status of Fixed/Incomplete/Invalid/Won't Fix in Launchpad - those are candidates for removal.
Note
Sometimes bugs are still New/Confirmed/Triaged/In Progress but have not had any hits in over 10 days. Those bugs should be re-assessed to see if they are now actually fixed or incomplete/invalid, marked as such and then remove the related query.
Running Queries Locally
You can execute an individual query locally and analyze the search results:
$ elastic-recheck-query queries/1331274.yaml
total hits: 133
build_status
100% FAILURE
build_name
48% check-grenade-dsvm
15% check-grenade-dsvm-partial-ncpu
13% gate-grenade-dsvm
9% check-grenade-dsvm-icehouse
9% check-grenade-dsvm-partial-ncpu-icehouse
build_branch
95% master
4% stable/icehouse
Notes
- The html generation will generate links that work with Kibana3's logstash.json dashboard. If you want the links to work properly on these generated files you will need to host a Kibana3 with that dashboard.
- View the OpenStack ElasticSearch cluster health here.
Future Work
- Move config files into a separate directory
- Make unit tests robust
- Add debug mode flag
- Expand gating testing
- Cleanup and document code better
- Add ability to check if any resolved bugs return
- Move away from polling ElasticSearch to discover if its ready or not
- Add nightly job to propose a patch to remove bug queries that return no hits -- Bug hasn't been seen in 2 weeks and must be closed