GeezerVison

OpenSearch is a multi-node Lucene implementation.

Logstash is a tool to collect, process, and forward events and log messages. Collection is accomplished via configurable input plugins including raw socket/packet communication, file tailing, and several message bus clients. Once an input plugin has collected data it can be processed by any number of filters which modify and annotate the event data. Finally logstash routes events to output plugins which can forward the events to a variety of external programs, local files, and several message bus implementations.

OpenSearch Dashboards is a browser-based analytics and search interface for OpenSearch.

Apache Kafka is a distributed streaming system. In our ELK stack Kafka buffers the stream of log messages produced by rsyslog (on behalf of applications) for consumption by Logstash. Nothing should output logs to logstash directly, logs should always be sent by way of Kafka.

Rsyslog is the "rocket-fast system for log processing". In our ELK stack rsyslog is used as the host "log agent". Rsyslog ingests log messages in various formats and from varying protocols, normalizes them and outputs to Kafka.

Support of logs shipped directly from application to Logstash has been deprecated.

Please see Logstash/Interface for details regarding long-term supported log shipping interfaces.

• EventLogging (of program-defined events with schemas), despite its name, uses a different pipeline.
• Varnish logs of the billions of pageviews of WMF wikis would require a lot more hardware. Instead we use Kafka to feed web requests into Hadoop. A notable exception to this rule: varnish user-facing errors (HTTP status 500-599) are sent to logstash to make debugging easier.
• While most of MediaWiki logs go to Logstash (and mwlog files), a few channels go exclusively to mwlog files. You can check which ones, via $wmgMonologChannels in InitialiseSettings.php.

When in the process of writing new logstash filters, take a look at what's existing already in puppet. Each filter must be tested to avoid regressions, we are using logstash filter verifier and existing tests can be found in the tests/ directory. To write tests or run existing tests you will need logstash-filter-verifier and logstash installed locally, or you can use docker/podman and the puppet repository:

# From the base dir of operations/puppet.
$ cd modules/profile/files/logstash/

# The Makefile recognizes if one of podman or docker is installed
# and then it uses it.
$ make test-local
/usr/bin/docker run --rm --workdir /src -v $(pwd):/src:Z -v $(pwd)/templates:/etc/logstash/templates:Z -v $(pwd)/filter_scripts:/etc/logstash/filter_scripts:Z --entrypoint make docker-registry.wikimedia.org/releng/logstash-filter-verifier:latest
logstash-filter-verifier --diff-command="diff -u --color=always" --sockets tests/ filters/*.conf
Use Unix domain sockets.
[...cut...]

Each filter has a corresponding test after its name in tests/. Within the test file the fields map lists the fields common to all tests and are used to trigger a specific filter's "if" conditions. The ignore key usually contains only @timestamp since that field is bound to change across invocations and can be safely ignored. The remainder of a test file is a list of testcases in the form of input/expected pairs. For "input" it is recommended to use yaml > to include verbatim JSON, whereas "expected" is usually yaml, although it can be also verbatim JSON if more convenient.

Please see Logstash/Interface#Tailing Log Files.

https://logstash.wikimedia.org

Log into https://logstash.wikimedia.org/ with your Wikimedia Developer account (sometimes known as your "LDAP" account).

Access to Logstash is restricted to members that are in one of the these LDAP groups, you may request access through the IDM tool, https://idm.wikimedia.org.

• logstash-access (for Wikimedia staff which needs Logstash)
• nda (for volunteers, soon this group will also migrate to logstash-access)
• ops (for SRE).

The cluster contains two types of nodes, configured by Puppet.

• role::logging::opensearch::collector manages the Logstash "collector" instances. These run Logstash, an OpenSearch indexing node, and an Apache vhost serving OpenSearch Dashboards. The Apache vhosts perform LDAP-based authentication to restrict access to the potentially sensitive log information.
• role::logging::opensearch::data configures an OpenSearch data node providing storage for log data.
• role::kafka::logging configures a Kafka broker for producers to publish log data to and for Logstash to consume from. This is a buffering layer to absorb log spikes and queue log events when maintenance is being performed on the logging cluster.

In October of 2023, the Observability team decided to rename the Logstash cluster to better align with the role(s) of each host. The previous naming convention, regardless of role or configuration, assigned each node the hostname logstash. This legacy naming convention required operators to know by the host ID (1001, et. al.) the role, function, and class.

The new naming convention selection criteria included:

• a concise name that could be expanded as needed
• indicated the difference between stateful and stateless components
• did not specify the underlying technology deployed
• an incremental improvement over role assignment based on node id

The 10-2023 Logging Cluster Naming Convention:

• logging-hd - OpenSearch data node: HDD Class
• logging-sd - OpenSearch data node: SSD Class
• logging-fe - OpenSearch API (logs-api), OpenSearch Dashboards (kibana7), Logstash Collector node

The "misc" Varnish cluster is being used to provide ssl termination and load balancing support for the Kibana application.

https://beta-logs.wmcloud.org/

Credentials for Beta's Logstash can be found on officewiki, or by connecting to deploy.svc.deployment-prep.eqiad1.wikimedia.cloud (points to a deployment server) and reading /root/secrets.txt. Unlike production services, Beta Cluster may not use Developer accounts (LDAP) for authentication.

$ ssh deploy.svc.deployment-prep.eqiad1.wikimedia.cloud -- sudo cat /root/secrets.txt
service: https://beta-logs.wmcloud.org
user: ************
password: ************

The security group kafka-logging must allow ingress from the logging collector on port 9093.

When commissioning a new logging collector, the certificate authority keystore (/etc/ssl/localcerts/wmf-java-cacerts) must be manually copied onto the new logging collector otherwise logstash will not start: (File does not exist or cannot be opened /etc/ssl/localcerts/wmf-java-cacerts).

The build can be run on the production builder host. See README for up-to-date build steps.

Make sure logging events sent to the GELF input don't have a "type" or "_type" field set, or if set, that it contains the value "gelf". The gelf/Logstash config discards any events that have a different value set for "type" or "_type". The final "type" seen in OpenSearch/Dashboards will be take from the "facility" element of the original GELF packet. The application sending the log data to Logstash should set "facility" to a reasonably unique value that identifies your application.

Log volume for a given type + channel + level + normalized message combination is throttled to 5000 / 5 min / Logstash collector (see 75-filter_throttle.conf). As of 2025 this means 30K events per 5 minutes (100/sec).

For a host of reasons it might happen that there's a buildup of messages on Kafka. For example:

OpenSearch is refusing to index messages, thus Logstash can't consume properly from Kafka.

The reason for index failure is usually conflicting fields, see also bug T150106 for a detailed discussion of the problem. The solution is to find what programs are generating the conflicts and drop them on Logstash accordingly, see also bug T228089

See the dead letter queue (DLQ) in logstash. We save a maximum of two days of dead-letters to assist with debugging indexing issues.

The warning shows up next to fields for which opensearch dashboards does not have mappings yet. To refresh the fields mapping navigate to index patterns, select the index pattern for your logs (likely logstash-* or ecs-*) then hit the "loop" or "refresh" icon on the top right.

After merging your configuration change Puppet will automatically restart Logstash. To force this to run:

 cumin -b1 -s60 'O:logging::opensearch::collector' 'run-puppet-agent -q'

 sudo /usr/share/logstash/bin/logstash --config.test_and_exit -f <myfile>

Have a look at the Dead Letter Queue Dashboard. The original message that caused the error is in the log.original field.

We're alerting on errors that Logstash gets from OpenSearch whenever there's an "indexing conflict" between fields of the same index (see also bug T236343). The reason usually is because two applications send logs with the same field name but two different types, e.g. response will be sent as a string in one case but as nested object in another. bug T239458 is a good example of this, where different parts of mediawiki send logs formatted in a different way.

This alert is based on the incoming logs per second indexed by OpenSearch. During normal operation there is a baseline of ~1k logs/s (July 2020) and anything significantly lower than that is an unexpected condition. Check the Logstash dashboard attached to alert for signs of root causes. Most likely Logstash has stopped sending logs to OpenSearch.

Occasionally producers will outpace Logstash's ingestion capabilities, most often with what's considered "log spam" (e.g. dumping whole request/response in debug logs). In these case one solution is to drop the offending logs from Logstash, and ideally the producer has already stopped spamming. The simplest such filter is installed before most/all other filters, matches a few fields and then drops the message:

filter {
  if [program] == "producer" and [nested][field] == "offending value" {
    drop {}
  }
}

See also this Gerrit change for a real-world example.

Logstash 5 locks up from time to time, causing UDP packet loss on the host it is running on. The fix in this case is to restart logstash.service on the host in question.

The storage drives on Logstash data are configured in an mdraid RAID0. OpenSearch handles data redundancy, so the rest of the cluster will absorb the impact of the downed node.

Once the disk is replaced, the RAID will have to be rebuilt:

First stop opensearch and disable puppet.

Copy disk partition layout from good disk to new disk

 sfdisk -d /dev/sdb | sfdisk /dev/sdi

determine md device mounted at /srv (/dev/md2 for example) and check mdstat

 cat /proc/mdstat

get array information. make a note of the remaining array members, we'll need this information when rebuilding

 mdadm --query --detail /dev/md2

stop and remove the raid0 array

 mdadm --stop /dev/md2 && mdadm --remove /dev/md2

remove traces of the previous array on the old partitions

 mdadm --zero-superblock /dev/sdb4
 mdadm --zero-superblock /dev/sdc4
 # ... etc
 mdadm --zero-superblock /dev/sdh4

create new raid0 array (WARNING: DISKS MAY BE DIFFERENT)

 mdadm --create --verbose /dev/md/2 --level=0 --raid-devices=8 /dev/sdb4 /dev/sdc4 /dev/sdd4 /dev/sde4 /dev/sdf4 /dev/sdg4 /dev/sdh4 /dev/sdi4

make filesystem

 mkfs.ext4 /dev/md2

workaround systemd mount management by commenting out the old array mount in fstab and issuing a daemon reload

 vim /etc/fstab
 systemctl daemon-reload

add mount point back in with new uuid and mount

 vim /etc/fstab
 mount /srv

check to make sure the disk is mounted and add new array definition to /etc/mdadm/mdadm.conf - also remove old definition

 mdadm --detail --scan
 vim /etc/mdadm/mdadm.conf

update initramfs

 update-initramfs -u

check other arrays for failed partitions and add partitions to them

 mdadm --manage /dev/md0 --add /dev/sdi2
 mdadm --manage /dev/md1 --add /dev/sdi3

make opensearch data directory

 mkdir /srv/opensearch && chown opensearch:opensearch /srv/opensearch

re-enable puppet and run puppet. OpenSearch should start up, join the cluster, and immediately start rebalancing shards.

Note: if an array is stuck PENDING with auto-read-write, it will either recover upon receiving its first write or you can manually invoke the resync with mdadm --readwrite /dev/mdX

From a single collector node, delete all .kibana indexes and restart opensearch-dashboards. Check that the restart created .kibana_1 and aliased it with .kibana. Fetch and unzip the backup and run:

 BACKUP_FILE=<myfile>.ndjson; curl -s -X POST http://localhost:5601/api/saved_objects/_import?createNewCopies=false -H "osd-xsrf: true" --form file=@$BACKUP_FILE > response.json

Check the response for problems and navigate into OpenSearch Dashboards to ensure expected saved objects are present.

Shard rebalancing operations are handled automatically by OpenSearch, however there are situations where shards cannot be assigned. Some of these possible reasons are:

• There is not enough space in the cluster to avoid exceeding a watermark.
  • This is often caused by the cluster storing too much data or a node has failed and the cluster is at reduced capacity. The solution is either to restore the failed node, reduce the number of replicas to free up space, or delete data to free up space.
• The index settings declare there should be more shards than there are number of nodes.
  • The solution is either to restore the failed node, reduce the number of replicas to free up space, or delete data to free up space.
• There is no host capable of accepting the shard due to label filters.
  • Ensure the cluster has a node labeled with the same value the index is expecting.

"RED" status: Occurs when OpenSearch reports there are unassigned primary shards. From the cluster perspective, this is a data-loss event. Usually, the only action required is to restore a node that has a full, uncorrupted, copy of the under-replicated shard.

"YELLOW" status: Occurs when OpenSearch reports there are unassigned replica shards.

"GREEN" status: Cluster is operating normally.

When hosts need to be rebooted (for ex. Kernel/firmware upgrades) it's necessary to follow this process to reduce the risk of data loss and to reduce the number of operations like leader elections.

There are two main types of Logstash hosts, the storage nodes and collector nodes, they require a different reboot process. The Logstash Grafana dashboard is a useful resource for monitoring node status including shard synchronization.

This process (including the logstash collectors) is now automated by the sre.opensearch.roll-restart-reboot cookbook!

The OpenSearch cat API provides a simple way to extract general statistics about log storage, e.g. total logs and bytes (not including replication)

 logstash1010:~$ curl -s 'localhost:9200/_cat/indices?v&bytes=b' | awk '/logstash-/ {b+=$10; d+=$7} END {print d; print b}'

Or logs per day (change $3 to $7 to get bytes sans replication)

 logstash1010:~$ curl -s 'localhost:9200/_cat/indices?v&bytes=b' | awk '/logstash-/ { gsub(/logstash-[^0-9]*/, "", $3); sum[$3] += $7 } END { for (i in sum) print i, sum[i] }' | sort

Or logs per month:

 logstash1010:~$ curl -s 'localhost:9200/_cat/indices?v&bytes=b' | awk '/logstash-/ { gsub(/logstash-[^0-9]*/, "", $3); gsub(/\.[0-9][0-9]$/, "", $3); sum[$3] += $7 } END { for (i in sum) print i, sum[i] }' | sort

See Logstash/Extended Retention