Troubleshooting Elasticsearch

DETAILS: Tier: Premium, Ultimate Offering: Self-managed, GitLab Dedicated

Use the following information to troubleshoot Elasticsearch issues.

Set configurations in the Rails console

List attributes

To list all available attributes:

Open the Rails console (sudo gitlab-rails console).
Run the following command:

ApplicationSetting.last.attributes

The output contains all the settings available in Elasticsearch integration, such as elasticsearch_indexing, elasticsearch_url, elasticsearch_replicas, and elasticsearch_pause_indexing.

Set attributes

To set an Elasticsearch integration setting, run a command like:

ApplicationSetting.last.update(elasticsearch_url: '<your ES URL and port>')

#or

ApplicationSetting.last.update(elasticsearch_indexing: false)

Get attributes

To check if the settings have been set in Elasticsearch integration or in the Rails console, run a command like:

Gitlab::CurrentSettings.elasticsearch_url

#or

Gitlab::CurrentSettings.elasticsearch_indexing

Change the password

To change the Elasticsearch password, run the following commands:

es_url = Gitlab::CurrentSettings.current_application_settings

# Confirm the current Elasticsearch URL
es_url.elasticsearch_url

# Set the Elasticsearch URL
es_url.elasticsearch_url = "http://<username>:<password>@your.es.host:<port>"

# Save the change
es_url.save!

View logs

One of the most valuable tools for identifying issues with the Elasticsearch integration are logs. The most relevant logs for this integration are:

sidekiq.log - All of the indexing happens in Sidekiq, so much of the relevant logs for the Elasticsearch integration can be found in this file.
elasticsearch.log - There are additional logs specific to Elasticsearch that are sent to this file that may contain useful diagnostic information about searching, indexing or migrations.

Here are some common pitfalls and how to overcome them.

Common terminology

Lucene: A full-text search library written in Java.
Near real time (NRT): Refers to the slight latency from the time to index a document to the time when it becomes searchable.
Cluster: A collection of one or more nodes that work together to hold all the data, providing indexing and search capabilities.
Node: A single server that works as part of a cluster.
Index: A collection of documents that have somewhat similar characteristics.
Document: A basic unit of information that can be indexed.
Shards: Fully-functional and independent subdivisions of indices. Each shard is actually a Lucene index.
Replicas: Failover mechanisms that duplicate indices.

How can you verify that your GitLab instance is using Elasticsearch?

There are a couple of ways to achieve that:

When you perform a search, in the upper-right corner of the search results page, Advanced search is enabled is displayed. This is always correctly identifying whether the current project/namespace being searched is using Elasticsearch.

From the Admin area under Settings > Search check that the advanced search settings are checked.

Those same settings there can be obtained from the Rails console if necessary:

::Gitlab::CurrentSettings.elasticsearch_search?         # Whether or not searches will use Elasticsearch
::Gitlab::CurrentSettings.elasticsearch_indexing?       # Whether or not content will be indexed in Elasticsearch
::Gitlab::CurrentSettings.elasticsearch_limit_indexing? # Whether or not Elasticsearch is limited only to certain projects/namespaces

Confirm searches use Elasticsearch by accessing the rails console and running the following commands:
```
u = User.find_by_email('email_of_user_doing_search')
s = SearchService.new(u, {:search => 'search_term'})
pp s.search_objects.class
```
The output from the last command is the key here. If it shows:
- ActiveRecord::Relation, it is not using Elasticsearch.
- Kaminari::PaginatableArray, it is using Elasticsearch.

If Elasticsearch is limited to specific namespaces and you need to know if Elasticsearch is being used for a specific project or namespace, you can use the Rails console:

::Gitlab::CurrentSettings.search_using_elasticsearch?(scope: Namespace.find_by_full_path("/my-namespace"))
::Gitlab::CurrentSettings.search_using_elasticsearch?(scope: Project.find_by_full_path("/my-namespace/my-project"))

Troubleshooting access

User: anonymous is not authorized to perform: es:ESHttpGet

When using a domain level access policy with AWS OpenSearch or Elasticsearch, the AWS role is not assigned to the correct GitLab nodes. The GitLab Rails and Sidekiq nodes require permission to communicate with the search cluster.

User: anonymous is not authorized to perform: es:ESHttpGet because no resource-based policy allows the es:ESHttpGet
action

To fix this, ensure the AWS role is assigned to the correct GitLab nodes.

Credential should be scoped to a valid region

When using AWS authorization with Advanced search, the region must be valid.

No permissions for `[indices:data/write/bulk]`

When using fine-grained access control with an IAM role or a role created using AWS OpenSearch Dashboards, you might encounter the following error:

{
  "error": {
    "root_cause": [
      {
        "type": "security_exception",
        "reason": "no permissions for [indices:data/write/bulk] and User [name=arn:aws:iam::xxx:role/INSERT_ROLE_NAME_HERE, backend_roles=[arn:aws:iam::xxx:role/INSERT_ROLE_NAME_HERE], requestedTenant=null]"
      }
    ],
    "type": "security_exception",
    "reason": "no permissions for [indices:data/write/bulk] and User [name=arn:aws:iam::xxx:role/INSERT_ROLE_NAME_HERE, backend_roles=[arn:aws:iam::xxx:role/INSERT_ROLE_NAME_HERE], requestedTenant=null]"
  },
  "status": 403
}

To fix this, you need to map the roles to users in the AWS OpenSearch Dashboards.

Create additional master users in AWS OpenSearch Service

You can set a master user when you create a domain. With this user, you can create additional master users. For more information, see the AWS documentation.

To create users and roles with permissions and map users to roles, see the OpenSearch documentation. You must include the following permissions in the role:

{
  "cluster_permissions": [
    "cluster_composite_ops",
    "cluster_monitor"
  ],
  "index_permissions": [
    {
      "index_patterns": [
        "gitlab*"
      ],
      "allowed_actions": [
        "data_access",
        "manage_aliases",
        "search",
        "create_index",
        "delete",
        "manage"
      ]
    },
    {
      "index_patterns": [
        "*"
      ],
      "allowed_actions": [
        "indices:admin/aliases/get",
        "indices:monitor/stats"
      ]
    }
  ]
}

Troubleshooting indexing

Troubleshooting indexing issues can be tricky. It can pretty quickly go to either GitLab support or your Elasticsearch administrator.

The best place to start is to determine if the issue is with creating an empty index. If it is, check on the Elasticsearch side to determine if the gitlab-production (the name for the GitLab index) exists. If it exists, manually delete it on the Elasticsearch side and attempt to recreate it from the recreate_index Rake task.

If you still encounter issues, try creating an index manually on the Elasticsearch instance. The details of the index aren't important here, as we want to test if indices can be made. If the indices:

Cannot be made, speak with your Elasticsearch administrator.
Can be made, Escalate this to GitLab support.

If the issue is not with creating an empty index, the next step is to check for errors during the indexing of projects. If errors do occur, they stem from either the indexing:

On the GitLab side. You need to rectify those. If they are not something you are familiar with, contact GitLab support for guidance.
Within the Elasticsearch instance itself. See if the error is documented and has a fix. If not, speak with your Elasticsearch administrator.

If the indexing process does not present errors, check the status of the indexed projects. You can do this via the following Rake tasks:

sudo gitlab-rake gitlab:elastic:index_projects_status (shows the overall status)
sudo gitlab-rake gitlab:elastic:projects_not_indexed (shows specific projects that are not indexed)

If:

Everything is showing at 100%, escalate to GitLab support. This could be a potential bug/issue.
You do see something not at 100%, attempt to reindex that project. To do this, run sudo gitlab-rake gitlab:elastic:index_projects ID_FROM=<project ID> ID_TO=<project ID>.

If reindexing the project shows:

Errors on the GitLab side, escalate those to GitLab support.
Elasticsearch errors or doesn't present any errors at all, reach out to your Elasticsearch administrator to check the instance.

You updated GitLab and now you can't find anything

We continuously make updates to our indexing strategies and aim to support newer versions of Elasticsearch. When indexing changes are made, it may be necessary for you to reindex after updating GitLab.

No search results in the UI after indexing all repositories

Make sure you indexed all the database data.

If there aren't any results (hits) in the UI search, check if you are seeing the same results via the rails console (sudo gitlab-rails console):

u = User.find_by_username('your-username')
s = SearchService.new(u, {:search => 'search_term', :scope => 'blobs'})
pp s.search_objects.to_a

Beyond that, check via the Elasticsearch Search API to see if the data shows up on the Elasticsearch side:

curl --request GET <elasticsearch_server_ip>:9200/gitlab-production/_search?q=<search_term>

More complex Elasticsearch API calls are also possible.

If the results:

Sync up, check that you are using supported syntax. Advanced search does not support exact substring matching.
Do not match up, this indicates a problem with the documents generated from the project. It is best to re-index that project.

NOTE: The above instructions are not to be used for scenarios that only index a subset of namespaces.

See Elasticsearch Index Scopes for more information on searching for specific types of data.

All repositories indexed, but no results after switching Elasticsearch servers

You must re-run all the Rake tasks to reindex the database, repositories, and wikis.

There are some projects that weren't indexed, but you don't know which ones

You can run sudo gitlab-rake gitlab:elastic:projects_not_indexed to display projects that aren't indexed.

No new data is added to the Elasticsearch index when you push code

NOTE: This was fixed in GitLab 13.2 and the Rake task is not available for versions greater than that.

When performing the initial indexing of blobs, we lock all projects until the project finishes indexing. It could happen that an error during the process causes one or multiple projects to remain locked. To unlock them, run:

sudo gitlab-rake gitlab:elastic:clear_locked_projects

Indexing fails with `error: elastic: Error 429 (Too Many Requests)`

If ElasticCommitIndexerWorker Sidekiq workers are failing with this error during indexing, it usually means that Elasticsearch is unable to keep up with the concurrency of indexing request. To address change the following settings:

To decrease the indexing throughput you can decrease Bulk request concurrency (see Advanced search settings). This is set to 10 by default, but you change it to as low as 1 to reduce the number of concurrent indexing operations.
If changing Bulk request concurrency didn't help, you can use the routing rules option to limit indexing jobs only to specific Sidekiq nodes, which should reduce the number of indexing requests.

Indexing is very slow or fails with `rejected execution of coordinating operation` messages

Bulk requests getting rejected by the Elasticsearch nodes are likely due to load and lack of available memory. Ensure that your Elasticsearch cluster meets the system requirements and has enough resources to perform bulk operations. See also the error "429 (Too Many Requests)".

Indexing fails with `strict_dynamic_mapping_exception`

Indexing might fail if all advanced search migrations were not finished before doing a major upgrade. A large Sidekiq backlog might accompany this error. To fix the indexing failures, you must re-index the database, repositories, and wikis.

Pause indexing so Sidekiq can catch up:

sudo gitlab-rake gitlab:elastic:pause_indexing

Recreate the index from scratch.

Resume indexing:

sudo gitlab-rake gitlab:elastic:resume_indexing

Indexing keeps pausing with `elasticsearch_pause_indexing setting is enabled`

You might notice that new data is not being detected when you run a search.

This error occurs when that new data is not being indexed properly.

To resolve this error, reindex your data.

However, when reindexing, you might get an error where the indexing process keeps pausing, and the Elasticsearch logs show the following:

"message":"elasticsearch_pause_indexing setting is enabled. Job was added to the waiting queue"

If reindexing does not resolve this issue, and you did not pause the indexing process manually, this error might be happening because two GitLab instances share one Elasticsearch cluster.

To resolve this error, disconnect one of the GitLab instances from using the Elasticsearch cluster.

For more information, see issue 3421.

Last resort to recreate an index

There may be cases where somehow data never got indexed and it's not in the queue, or the index is somehow in a state where migrations just cannot proceed. It is always best to try to troubleshoot the root cause of the problem by viewing the logs.

As a last resort, you can recreate the index from scratch. For small GitLab installations, recreating the index can be a quick way to resolve some issues. For large GitLab installations, however, this method might take a very long time. Your index does not show correct search results until the indexing is complete. You might want to clear the Search with Elasticsearch enabled checkbox while the indexing is running.

If you are sure you've read the above caveats and want to proceed, then you should run the following Rake task to recreate the entire index from scratch.

::Tabs

:::TabTitle Linux package (Omnibus)

# WARNING: DO NOT RUN THIS UNTIL YOU READ THE DESCRIPTION ABOVE
sudo gitlab-rake gitlab:elastic:index

:::TabTitle Self-compiled (source)

# WARNING: DO NOT RUN THIS UNTIL YOU READ THE DESCRIPTION ABOVE
cd /home/git/gitlab
sudo -u git -H bundle exec rake gitlab:elastic:index

::EndTabs

Troubleshooting performance

Troubleshooting performance can be difficult on Elasticsearch. There is a ton of tuning that can be done, but the majority of this falls on shoulders of a skilled Elasticsearch administrator.

Generally speaking, ensure:

The Elasticsearch server is not running on the same node as GitLab.
The Elasticsearch server have enough RAM and CPU cores.
That sharding is being used.

Going into some more detail here, if Elasticsearch is running on the same server as GitLab, resource contention is very likely to occur. Ideally, Elasticsearch, which requires ample resources, should be running on its own server (maybe coupled with Logstash and Kibana).

When it comes to Elasticsearch, RAM is the key resource. Elasticsearch themselves recommend:

At least 8 GB of RAM for a non-production instance.
At least 16 GB of RAM for a production instance.
Ideally, 64 GB of RAM.

For CPU, Elasticsearch recommends at least 2 CPU cores, but Elasticsearch states common setups use up to 8 cores. For more details on server specs, check out the Elasticsearch hardware guide.

Beyond the obvious, sharding comes into play. Sharding is a core part of Elasticsearch. It allows for horizontal scaling of indices, which is helpful when you are dealing with a large amount of data.

With the way GitLab does indexing, there is a huge amount of documents being indexed. By using sharding, you can speed up the ability of Elasticsearch to locate data because each shard is a Lucene index.

If you are not using sharding, you are likely to hit issues when you start using Elasticsearch in a production environment.

An index with only one shard has no scale factor and is likely to encounter issues when called upon with some frequency. See the Elasticsearch documentation on capacity planning.

The easiest way to determine if sharding is in use is to check the output of the Elasticsearch Health API:

Red means the cluster is down.
Yellow means it is up with no sharding/replication.
Green means it is healthy (up, sharding, replicating).

For production use, it should always be green.

Beyond these steps, you get into some of the more complicated things to check, such as merges and caching. These can get complicated and it takes some time to learn them, so it is best to escalate/pair with an Elasticsearch expert if you need to dig further into these.

Feel free to reach out to GitLab support, but this is likely to be something a skilled Elasticsearch administrator has more experience with.

Slow initial indexing

The more data your GitLab instance has, the longer the indexing takes. You can estimate cluster size with the Rake task sudo gitlab-rake gitlab:elastic:estimate_cluster_size.

For code documents

Ensure you have enough Sidekiq nodes and processes to efficiently index code, commits, and wikis. If your initial indexing is slow, consider dedicated Sidekiq nodes or processes.

For non-code documents

If the initial indexing is slow but Sidekiq has enough nodes and processes, you can adjust advanced search worker settings in GitLab. For Requeue indexing workers, the default value is false. For Number of shards for non-code indexing, the default value is 2. These settings limit indexing to 2000 documents per minute.

To adjust worker settings:

On the left sidebar, at the bottom, select Admin.
Select Settings > Search.
Expand Advanced Search.
Select the Requeue indexing workers checkbox.
In the Number of shards for non-code indexing text box, enter a value higher than 2.
Select Save changes.

Issues with migrations

Ensure you've read about Elasticsearch Migrations.

If there is a halted migration and your elasticsearch.log file contain errors, this could potentially be a bug/issue. Escalate to GitLab support if retrying migrations does not succeed.

Error: `Can't specify parent if no parent field has been configured`

If you enabled Elasticsearch before GitLab 8.12 and have not rebuilt indices, you get exceptions in lots of different cases:

Elasticsearch::Transport::Transport::Errors::BadRequest([400] {
    "error": {
        "root_cause": [{
            "type": "illegal_argument_exception",
            "reason": "Can't specify parent if no parent field has been configured"
        }],
        "type": "illegal_argument_exception",
        "reason": "Can't specify parent if no parent field has been configured"
    },
    "status": 400
}):

This is because we changed the index mapping in GitLab 8.12 and the old indices should be removed and built from scratch again, see details in the update guide.

Error: `Elasticsearch::Transport::Transport::Errors::BadRequest`

If you have this exception (just like in the case above but the actual message is different), check that you have the correct Elasticsearch version and you met the other requirements. There is also an easy way to check it automatically with sudo gitlab-rake gitlab:check command.

Error: `Elasticsearch::Transport::Transport::Errors::RequestEntityTooLarge`

[413] {"Message":"Request size exceeded 10485760 bytes"}

This exception is seen when your Elasticsearch cluster is configured to reject requests above a certain size (10 MiB in this case). This corresponds to the http.max_content_length setting in elasticsearch.yml. Increase it to a larger size and restart your Elasticsearch cluster.

AWS has network limits on the maximum size of HTTP request payloads based on the size of the underlying instance. Set the maximum bulk request size to a value lower than 10 MiB.

Error: `Faraday::TimeoutError (execution expired)`

When you use a proxy, set a custom gitlab_rails['env'] environment variable named no_proxy with the IP address of your Elasticsearch host.

My single node Elasticsearch cluster status never goes from `yellow` to `green`

For a single node Elasticsearch cluster the functional cluster health status is yellow (never green) because the primary shard is allocated but replicas cannot be as there is no other node to which Elasticsearch can assign a replica. This also applies if you are using the Amazon OpenSearch service.

WARNING: Setting the number of replicas to 0 is discouraged (this is not allowed in the GitLab Elasticsearch Integration menu). If you are planning to add more Elasticsearch nodes (for a total of more than 1 Elasticsearch) the number of replicas needs to be set to an integer value larger than 0. Failure to do so results in lack of redundancy (losing one node corrupts the index).

If you have a hard requirement to have a green status for your single node Elasticsearch cluster, make sure you understand the risks outlined in the previous paragraph and then run the following query to set the number of replicas to 0(the cluster no longer tries to create any shard replicas):

curl --request PUT localhost:9200/gitlab-production/_settings --header 'Content-Type: application/json' \
     --data '{
       "index" : {
         "number_of_replicas" : 0
       }
     }'

Error: `health check timeout: no Elasticsearch node available`

If you're getting a health check timeout: no Elasticsearch node available error in Sidekiq during the indexing process:

Gitlab::Elastic::Indexer::Error: time="2020-01-23T09:13:00Z" level=fatal msg="health check timeout: no Elasticsearch node available"

You probably have not used either http:// or https:// as part of your value in the "URL" field of the Elasticsearch Integration Menu. Make sure you are using either http:// or https:// in this field as the Elasticsearch client for Go that we are using needs the prefix for the URL to be accepted as valid. After you have corrected the formatting of the URL, delete the index (via the dedicated Rake task) and reindex the content of your instance.

My Elasticsearch cluster has a plugin and the integration is not working

Certain 3rd party plugins may introduce bugs in your cluster or for whatever reason may be incompatible with our integration. You should try disabling plugins so you can rule out the possibility that the plugin is causing the problem.

Elasticsearch `code_analyzer` doesn't account for all code cases

The code_analyzer pattern and filter configuration is being evaluated for improvement. We have fixed most edge cases that were not returning expected search results due to our pattern and filter configuration.

Improvements to the code_analyzer pattern and filters are being discussed in epic 3621.

Some binary files may not be searchable by name

In GitLab 13.9, a change was made where binary filenames are being indexed. However, without indexing all projects' data from scratch, only binary files that are added or updated after the GitLab 13.9 release are searchable.

How does advanced search handle private projects?

Advanced search stores all the projects in the same Elasticsearch indices, however, searches only surface results that can be viewed by the user. Advanced search honors all permission checks in the application by filtering out projects that a user does not have access to at search time.

Elasticsearch workers overload Sidekiq

In some cases, Elasticsearch cannot connect to GitLab anymore because:

The Elasticsearch password has been updated on one side only (Unauthorized [401] ... unable to authenticate user errors).
A firewall or network issue impairs connectivity (Failed to open TCP connection to <ip>:9200 errors).

These errors are logged in gitlab-rails/elasticsearch.log. To retrieve the errors, use jq:

$ jq --raw-output 'select(.severity == "ERROR") | [.error_class, .error_message] | @tsv' \
    gitlab-rails/elasticsearch.log |
  sort | uniq -c

Elastic workers and Sidekiq jobs could also appear much more often because Elasticsearch frequently attempts to reindex if a previous job fails. You can use fast-stats or jq to count workers in the Sidekiq logs:

$ fast-stats --print-fields=count,score sidekiq/current
WORKER                            COUNT   SCORE
ElasticIndexBulkCronWorker          234  123456
ElasticIndexInitialBulkCronWorker   345   12345
Some::OtherWorker                    12     123
...

$ jq '.class' sidekiq/current | sort | uniq -c | sort -nr
 234 "ElasticIndexInitialBulkCronWorker"
 345 "ElasticIndexBulkCronWorker"
  12 "Some::OtherWorker"
...

In this case, free -m on the overloaded GitLab node would also show unexpectedly high buff/cache usage.

Error: `Couldn't load task status`

When you reindex, you might get a Couldn't load task status error. A sliceId must be greater than 0 but was [-1] error might also appear on the Elasticsearch host. As a workaround, consider reindexing from scratch or upgrading to GitLab 16.3.

For more information, see issue 422938.

Migration `BackfillProjectPermissionsInBlobs` has been halted in GitLab 15.11

In GitLab 15.11, it is possible for the BackfillProjectPermissionsInBlobs migration to be halted with the following error message in the elasticsearch.log:

migration has failed with NoMethodError:undefined method `<<' for nil:NilClass, no retries left

If BackfillProjectPermissionsInBlobs is the only halted migration, you can upgrade to the latest patch version of GitLab 16.0, which includes the fix. Otherwise, you can ignore the error as it will not affect the current functionality of advanced search.

`ElasticIndexInitialBulkCronWorker` and `ElasticIndexBulkCronWorker` jobs stuck in deduplication

In GitLab 16.5 and earlier, the ElasticIndexInitialBulkCronWorker and ElasticIndexBulkCronWorker jobs might get stuck in deduplication. This issue might prevent advanced search from properly indexing documents even after creating a new index. In GitLab 16.6, idempotent! was removed for bulk cron workers that perform indexing.

The Sidekiq log might have the following entries:

{"severity":"INFO","time":"2023-10-31T10:33:06.998Z","retry":0,"queue":"default","version":0,"queue_namespace":"cronjob","args":[],"class":"ElasticIndexInitialBulkCronWorker",
...
"idempotency_key":"resque:gitlab:duplicate:default:<value>","duplicate-of":"91e8673347d4dc84fbad5319","job_size_bytes":2,"pid":12047,"job_status":"deduplicated","message":"ElasticIndexInitialBulkCronWorker JID-5e1af9180d6e8f991fc773c6: deduplicated: until executing","deduplication.type":"until executing"}

To resolve this issue:

In a Rails console session, run this command:

idempotency_key = "<idempotency_key_from_log_entry>"
duplicate_key = "resque:gitlab:#{idempotency_key}:cookie:v2"
Gitlab::Redis::Queues.with { |c| c.del(duplicate_key) }

Replace <idempotency_key_from_log_entry> with the actual entry in your log.