[IMPALA-9241] Minicluster service status ambiguous when pids wrap around

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 3.4.0
• Fix Version/s: Impala 3.4.0
• Component/s: Infrastructure
• Labels: broken-build
• Target Version: Impala 3.4.0
• Epic Color: ghx-label-12
• Assignee: Joe McDonnell
• Reporter: Joe McDonnell
• Created: 13/Dec/19 01:08
• Updated: 03/Jan/20 18:43
• Resolved: 03/Jan/20 18:42

Description

In a recent test run, a large number of tests failed due to being unable to contact the Kudu master. Messages like:

 

query_test/test_acid.py:26: in <module>
    from tests.common.skip import (SkipIfHive2, SkipIfCatalogV2, SkipIfS3, SkipIfABFS,
common/skip.py:108: in <module>
    class SkipIfKudu:
common/skip.py:112: in SkipIfKudu
    get_kudu_master_flag("--use_hybrid_clock") == "false",
common/kudu_test_suite.py:59: in get_kudu_master_flag
    varz = get_kudu_master_webpage("varz")
common/kudu_test_suite.py:55: in get_kudu_master_webpage
    return requests.get(url).text
/home/user/Impala/infra/python/env/lib/python2.7/site-packages/requests/api.py:69: in get
    return request('get', url, params=params, **kwargs)
/home/user/Impala/infra/python/env/lib/python2.7/site-packages/requests/api.py:50: in request
    response = session.request(method=method, url=url, **kwargs)
/home/user/Impala/infra/python/env/lib/python2.7/site-packages/requests/sessions.py:465: in request
    resp = self.send(prep, **send_kwargs)
/home/user/Impala/infra/python/env/lib/python2.7/site-packages/requests/sessions.py:573: in send
    r = adapter.send(request, **kwargs)
/home/user/Impala/infra/python/env/lib/python2.7/site-packages/requests/adapters.py:415: in send
    raise ConnectionError(err, request=request)
E   ConnectionError: ('Connection aborted.', error(111, 'Connection refused'))

I checked the logs/cluster/cdh6-node-1/kudu/master directory, and there was a log file for the minicluster startup to do dataload, but not one for the restart of the minicluster at the end of dataload (https://github.com/apache/impala/blob/fc4a91cf8c87966a910106dded7e7eb8d215270a/testdata/bin/create-load-data.sh#L717).

 

Interestingly, two of the tablet servers did start up as part of that restart, so this was not a universal thing. In fact, quite a few tests ran fine.

My theory is that this could be due to stale PIDs. When starting up one of the services covered by testdata/cluster/admin, it calls the status function to see if the service is already running (https://github.com/apache/impala/blob/master/testdata/cluster/admin#L423):

 

if "$SCRIPT" status &>/dev/null; then  
  RUNNING=true
else
  RUNNING=false
fi

If it is already RUNNING, it skips the startup. The status call is common across the different services and reads the PID from a file and checks to see if that PID is still running:

 

 

function status {
  local PID=$(read_pid)
  if [[ -z $PID ]]; then
    echo Not started
    return 1
  fi
  if pid_exists $PID; then
    echo Running
  else
    echo Not Running
    return 1
  fi
}

However, it doesn't delete the pid file when it shuts down. If something happens to be running with that pid when we try to start up, it would think it is already running and fail to start up. Silently!

This could apply to kudu, hdfs, kms, yarn, etc. It does not apply to hbase, hive, sentry, ranger, as those do not use this framework.