Skip to content

Latest commit

 

History

History
204 lines (151 loc) · 9.47 KB

DEPLOYMENT.md

File metadata and controls

204 lines (151 loc) · 9.47 KB
Raw

Simple Deployment Procedure

In order to update the codebase in the vocms0241 virtual machine,

  • Please push your commits or open your PRs against "https://github.com/dmwm/cms-htcondor-es/tree/master".
  • In virtual machine, fetch: git fetch --all
  • Check the difference between local and remote to be sure: git diff origin/master master
  • Before rebase, please make sure if 12m period is finished:
    • No python spider_cms process: ps aux | grep "python spider_cms.py"
    • And tail -f log/spider_cms.log is like @@@ Total processing time:...
  • Finally rebase git pull --rebase origin master

General Overview

https://cmsmonit-docs.web.cern.ch/cms-htcondor-es/spider/

PRODUCTION DETAILS

What is running:

Crontab

*/12 * * * * /home/cmsjobmon/cms-htcondor-es/scripts/spider_cms_queues.sh
5-59/12 * * * * /home/cmsjobmon/cms-htcondor-es/scripts/spider_cms_history.sh
0 3 * * * /bin/bash "/home/cmsjobmon/cms-htcondor-es/scripts/cronAffiliation.sh"
# 0 * * * * /bin/bash "/home/cmsjobmon/scripts/sourcesCompare_cron.sh"

Users

  • cmsjobmon user is used to run the cron jobs.
  • logstash user is used to run the Logstash service.

VENV

  • Current virtual machine has python3.9 and before deployment, new virtual environment should be created with venv3_9 name.
  • Requirements should be installed before cron jobs are activated.
  • To make migrations easy, please use meaningful "venv" names:
cd /home/cmsjobmon/cms-htcondor-es && python3 -m venv venv3_X && ./venv3_X/bin/pip install --no-cache-dir -r requirements.txt

Required directories, files and secrets

Required CERN CA certificate: /etc/pki/tls/certs/CERN-bundle.pem

Used environment variables in scripts:

export SPIDER_WORKDIR="/home/cmsjobmon/cms-htcondor-es"
export AFFILIATION_DIR_LOCATION="$SPIDER_WORKDIR/.affiliation_dir.json"
Under HOME /home/cmsjobmon
file path description
SPIDER_WORKDIR $HOME/cms-htcondor-es MAIN
usercert.pem $HOME/.globus/usercert.pem used by Affiliation cron job to query CRIC and update affiliations
userkey.pem $HOME/.globus/userkey.pem used by Affiliation cron job to query CRIC and update affiliations
Under SPIDER_WORKDIR
file parametrized value description
amq_username $SPIDER_WORKDIR/etc/amq_username ActiveMQ user
amq_password $SPIDER_WORKDIR/etc/amq_password ActiveMQ password
.affiliation_dir.json $SPIDER_WORKDIR/.affiliation_dir.json stores affiliations of all users in JSON format
checkpoint.json $SPIDER_WORKDIR/checkpoint.json last update times of each history schedd
collectors.json $SPIDER_WORKDIR/etc/collectors.json Collectors that will be queried
es_conf.json $SPIDER_WORKDIR/etc/es_conf.json CMS os-cms OpenSearch host, username, password credentials
JobMonitoring.json $SPIDER_WORKDIR/JobMonitoring.json ClassAds to JSON format conversion schema
last_mappings.json $SPIDER_WORKDIR/last_mappings.json CMS os-cms OpenSearch index mapping
log, log_history, log_aff $SPIDER_WORKDIR/log* Log directories
venv3_9 $SPIDER_WORKDIR/venv3_9 used venv

Prod es_conf.json format

CONVENTIONS:

  • Host name ends with /os refers to OpenSearch otherwise it is behaved as ElasticSearch and port should be provided.
  • No need to provide port for OpenSearch.
  • Supports multiple clusters and send data to all of them without querying schedds again.
  • Do not put https:// prefix.
[
  {
    "host": "es-cmsX(ElasticSearch).cern.ch",
    "port": 9203,
    "username": "user",
    "password": "pass"
  },
  {
    "host": "os-cmsY(OpenSearch).cern.ch/os",
    "username": "user",
    "password": "pass"
  }
]

Example scenario to send data to 2 OpenSearch instances:

[
  {"host": "os-cms1.cern.ch/os", "username": "user", "password": "pass"},
  {"host": "os-cms2.cern.ch/os", "username": "user", "password": "pass"}
]

Test

MIGRATION AND DISASTER RECOVERY

Migration

  • Prepare all the production requirements: directory, files, secrets and venv.
  • Take backup of $SPIDER_WORKDIR/checkpoint.json as $SPIDER_WORKDIR/checkpoint.json.back.
  • Set new crontab and let them start.
  • For LogStash and Filebeat migration, please follow documentation of: cms-htcondor-es/service-logstash

Disaster Recovery

Scenario-1: Data send wrongly

  • Rollback to previous commit, directory&file structure and venv.
  • Rollback to back-up checkpoint: mv checkpoint.json.back checkpoint.json.
  • Delete documents from os-cms.cern.ch cms-* indices with metadata.spider_git_hash='problematic_hash'.
  • Ask MONIT to delete documents from monit-opensearch.cern.ch monit_prod_condor_raw_metric* indices with data.metadata.spider_git_hash='problematic_hash'.
  • Inform and announce HDFS users to filter out metadata.spider_git_hash='problematic_hash' in their Spark jobs.
    • HDFS data deletion: it is not easy because of Flume data. It can be deleted after a couple of days later. Discuss with MONIT.
  • Data loss is inevitable(12m or more) for Condor Job Monitoring schedds Queue data (Running, Held, Idle, etc.).
  • No data loss for schedds History data (completed, etc.) because of checkpoint.json if required actions are taken immediately.

Scenario-2: not worked at all

  • Rollback to previous commit, directory&file structure and venv.
  • Rollback to back-up checkpoint: mv checkpoint.json.back checkpoint.json.
  • Data loss is inevitable(12m or more) for Condor Job Monitoring schedds Queue data (Running, Held, Idle, etc.).
  • No data loss for schedds History data (Completed, etc.) because of checkpoint.json if required actions are taken immediately.

Scenario-3: failed after some time later

In any case, data loss is inevitable for Condor Job Monitoring schedds Queue data (Running, Held, Idle, etc.)

  • If problem is noticed within 12 hours, no action needed. Just fix the problem and continue to run.
  • If problem is noticed later than 12 hours, set --history_query_max_n_minutes=12*60 parameter in spider_cms_history.sh to a reasonable time window to include last checkpoints of checkpoint.json to fill the schedds History results. Else, data loss will be in case for schedds History data. It is because of a safety measure: even if schedds have previous checkpoint time than history_query_max_n_minutes, their checkpoints will be set to this max time window. It can be tweaked in disaster scenarios.

Opensearch help utils

# Get indices
curl -s -XGET --negotiate -u $user:$pass https://os-cms1.cern.ch/os/_cat/indices | grep ceyhun | sort

# Delete one index
curl -s -XDELETE --negotiate -u $user:$pass https://os-cms1.cern.ch/os/cms-test-ceyhun-2023-02-21

# Deletion command(NO RUN, just echo) for multiple indices with regex
echo 'curl -XDELETE --negotiate -u $user:$pass' https://os-cms1.cern.ch/os/"$(
     curl -s -XGET --negotiate -u $user:$pass https://os-cms1.cern.ch/os/_cat/indices/cms-test-ceyhun* | sort | awk '{ORS=","; print $3}'
)"

Elasticsearch help utils

# Get indices
curl -s -XGET -u $user:$pass https://es-cms.cern.ch:9203/_cat/indices/ | grep ceyhun | sort

# Delete one indice
curl -s -XDELETE -u $user:$pass https://es-cms.cern.ch:9203/cms-test-ceyhun-2023-02-21

# Deletion command(NO RUN, just echo) for multiple indices with regex
echo 'curl -XDELETE -u $user:$pass' https://es-cms.cern.ch:9203/"$(
    curl -s -XGET -u $user:$pass https://es-cms.cern.ch:9203/_cat/indices/cms-test-ceyhun* | sort | awk '{ORS=","; print $3}')"