Advanced search development guidelines

This page includes information about developing and working with Elasticsearch.

Information on how to enable Elasticsearch and perform the initial indexing is in the Elasticsearch integration documentation.

Deep Dive

In June 2019, Mario de la Ossa hosted a Deep Dive (GitLab team members only: https://gitlab.com/gitlab-org/create-stage/-/issues/1) on the GitLab Elasticsearch integration to share his domain specific knowledge with anyone who may work in this part of the codebase in the future. You can find the recording on YouTube, and the slides on Google Slides and in PDF. Everything covered in this deep dive was accurate as of GitLab 12.0, and while specific details might have changed, it should still serve as a good introduction.

In August 2020, a second Deep Dive was hosted, focusing on GitLab-specific architecture for multi-indices support. The recording on YouTube and the slides are available. Everything covered in this deep dive was accurate as of GitLab 13.3.

Supported Versions

See Version Requirements.

Developers making significant changes to Elasticsearch queries should test their features against all our supported versions.

Setting up development environment

See the Elasticsearch GDK setup instructions

Helpful Rake tasks

gitlab:elastic:test:index_size: Tells you how much space the current index is using, as well as how many documents are in the index.
gitlab:elastic:test:index_size_change: Outputs index size, reindexes, and outputs index size again. Useful when testing improvements to indexing size.

Additionally, if you need large repositories or multiple forks for testing, please consider following these instructions

How does it work?

The Elasticsearch integration depends on an external indexer. We ship an indexer written in Go. The user must trigger the initial indexing via a Rake task but, after this is done, GitLab itself will trigger reindexing when required via after_ callbacks on create, update, and destroy that are inherited from /ee/app/models/concerns/elastic/application_versioned_search.rb.

After initial indexing is complete, create, update, and delete operations for all models except projects (see #207494) are tracked in a Redis ZSET. A regular sidekiq-cron ElasticIndexBulkCronWorker processes this queue, updating many Elasticsearch documents at a time with the Bulk Request API.

Search queries are generated by the concerns found in ee/app/models/concerns/elastic. These concerns are also in charge of access control, and have been a historic source of security bugs so please pay close attention to them!

Custom routing

Custom routing is used in Elasticsearch for document types that are associated with a project. The routing format is project_<project_id>. Routing is set during indexing and searching operations. Some of the benefits and tradeoffs to using custom routing are:

Project scoped searches are much faster.
Routing is not used if too many shards would be hit for global and group scoped searches.
Shard size imbalance might occur.

Existing analyzers and tokenizers

The following analyzers and tokenizers are defined in ee/lib/elastic/latest/config.rb.

Analyzers

`path_analyzer`

Used when indexing blobs' paths. Uses the path_tokenizer and the lowercase and asciifolding filters.

Please see the path_tokenizer explanation below for an example.

`sha_analyzer`

Used in blobs and commits. Uses the sha_tokenizer and the lowercase and asciifolding filters.

Please see the sha_tokenizer explanation later below for an example.

`code_analyzer`

Used when indexing a blob's filename and content. Uses the whitespace tokenizer and the word_delimiter_graph, lowercase, and asciifolding filters.

The whitespace tokenizer was selected to have more control over how tokens are split. For example the string Foo::bar(4) needs to generate tokens like Foo and bar(4) to be properly searched.

Please see the code filter for an explanation on how tokens are split.

NOTE: The Elasticsearch code_analyzer doesn't account for all code cases.

Tokenizers

`sha_tokenizer`

This is a custom tokenizer that uses the edgeNGram tokenizer to allow SHAs to be searchable by any sub-set of it (minimum of 5 chars).

Example:

240c29dc7e becomes:

240c2
240c29
240c29d
240c29dc
240c29dc7
240c29dc7e

`path_tokenizer`

This is a custom tokenizer that uses the path_hierarchy tokenizer with reverse: true to allow searches to find paths no matter how much or how little of the path is given as input.

Example:

'/some/path/application.js' becomes:

'/some/path/application.js'
'some/path/application.js'
'path/application.js'
'application.js'

Gotchas

Searches can have their own analyzers. Remember to check when editing analyzers.
Character filters (as opposed to token filters) always replace the original character. These filters can hinder exact searches.

Zero downtime reindexing with multiple indices

NOTE: This is not applicable yet as multiple indices functionality is not fully implemented.

Currently GitLab can only handle a single version of setting. Any setting/schema changes would require reindexing everything from scratch. Since reindexing can take a long time, this can cause search functionality downtime.

To avoid downtime, GitLab is working to support multiple indices that can function at the same time. Whenever the schema changes, the administrator will be able to create a new index and reindex to it, while searches continue to go to the older, stable index. Any data updates will be forwarded to both indices. Once the new index is ready, an administrator can mark it active, which will direct all searches to it, and remove the old index.

This is also helpful for migrating to new servers, for example, moving to/from AWS.

Currently we are on the process of migrating to this new design. Everything is hardwired to work with one single version for now.

Architecture

The traditional setup, provided by elasticsearch-rails, is to communicate through its internal proxy classes. Developers would write model-specific logic in a module for the model to include in (for example, SnippetsSearch). The __elasticsearch__ methods would return a proxy object, for example:

Issue.__elasticsearch__ returns an instance of Elasticsearch::Model::Proxy::ClassMethodsProxy
Issue.first.__elasticsearch__ returns an instance of Elasticsearch::Model::Proxy::InstanceMethodsProxy.

These proxy objects would talk to Elasticsearch server directly (see top half of the diagram).

In the planned new design, each model would have a pair of corresponding sub-classed proxy objects, in which model-specific logic is located. For example, Snippet would have SnippetClassProxy being a subclass of Elasticsearch::Model::Proxy::ClassMethodsProxy. Snippet would have SnippetInstanceProxy being a subclass of Elasticsearch::Model::Proxy::InstanceMethodsProxy.

__elasticsearch__ would represent another layer of proxy object, keeping track of multiple actual proxy objects. It would forward method calls to the appropriate index. For example:

model.__elasticsearch__.search would be forwarded to the one stable index, since it is a read operation.
model.__elasticsearch__.update_document would be forwarded to all indices, to keep all indices up-to-date.

The global configurations per version are now in the Elastic::(Version)::Config class. You can change mappings there.

Creating new version of schema

NOTE: This is not applicable yet as multiple indices functionality is not fully implemented.

Folders like ee/lib/elastic/v12p1 contain snapshots of search logic from different versions. To keep a continuous Git history, the latest version lives under ee/lib/elastic/latest, but its classes are aliased under an actual version (for example, ee/lib/elastic/v12p3). When referencing these classes, never use the Latest namespace directly, but use the actual version (for example, V12p3).

The version name basically follows the GitLab release version. If setting is changed in 12.3, we will create a new namespace called V12p3 (p stands for "point"). Raise an issue if there is a need to name a version differently.

If the current version is v12p1, and we need to create a new version for v12p3, the steps are as follows:

Copy the entire folder of v12p1 as v12p3
Change the namespace for files under v12p3 folder from V12p1 to V12p3 (which are still aliased to Latest)
Delete v12p1 folder
Copy the entire folder of latest as v12p1
Change the namespace for files under v12p1 folder from Latest to V12p1
Make changes to files under the latest folder as needed

Performance Monitoring

Prometheus

GitLab exports Prometheus metrics relating to the number of requests and timing for all web/API requests and Sidekiq jobs, which can help diagnose performance trends and compare how Elasticsearch timing is impacting overall performance relative to the time spent doing other things.

Indexing queues

GitLab also exports Prometheus metrics for indexing queues, which can help diagnose performance bottlenecks and determine whether or not your GitLab instance or Elasticsearch server can keep up with the volume of updates.

Logs

All of the indexing happens in Sidekiq, so much of the relevant logs for the Elasticsearch integration can be found in sidekiq.log. In particular, all Sidekiq workers that make requests to Elasticsearch in any way will log the number of requests and time taken querying/writing to Elasticsearch. This can be useful to understand whether or not your cluster is keeping up with indexing.

Searching Elasticsearch is done via ordinary web workers handling requests. Any requests to load a page or make an API request, which then make requests to Elasticsearch, will log the number of requests and the time taken to production_json.log. These logs will also include the time spent on Database and Gitaly requests, which may help to diagnose which part of the search is performing poorly.

There are additional logs specific to Elasticsearch that are sent to elasticsearch.log that may contain information to help diagnose performance issues.

Performance Bar

Elasticsearch requests will be displayed in the Performance Bar, which can be used both locally in development and on any deployed GitLab instance to diagnose poor search performance. This will show the exact queries being made, which is useful to diagnose why a search might be slow.

Correlation ID and `X-Opaque-Id`

Our correlation ID is forwarded by all requests from Rails to Elasticsearch as the X-Opaque-Id header which allows us to track any tasks in the cluster back the request in GitLab.

Troubleshooting

Getting `flood stage disk watermark [95%] exceeded`

You might get an error such as

[2018-10-31T15:54:19,762][WARN ][o.e.c.r.a.DiskThresholdMonitor] [pval5Ct]
   flood stage disk watermark [95%] exceeded on
   [pval5Ct7SieH90t5MykM5w][pval5Ct][/usr/local/var/lib/elasticsearch/nodes/0] free: 56.2gb[3%],
   all indices on this node will be marked read-only

This is because you've exceeded the disk space threshold - it thinks you don't have enough disk space left, based on the default 95% threshold.

In addition, the read_only_allow_delete setting will be set to true. It will block indexing, forcemerge, etc

curl "http://localhost:9200/gitlab-development/_settings?pretty"

Add this to your elasticsearch.yml file:

# turn off the disk allocator
cluster.routing.allocation.disk.threshold_enabled: false

# set your own limits
cluster.routing.allocation.disk.threshold_enabled: true
cluster.routing.allocation.disk.watermark.flood_stage: 5gb   # ES 6.x only
cluster.routing.allocation.disk.watermark.low: 15gb
cluster.routing.allocation.disk.watermark.high: 10gb

Restart Elasticsearch, and the read_only_allow_delete will clear on its own.

from "Disk-based Shard Allocation | Elasticsearch Reference" 5.6 and 6.x

Disaster recovery/data loss/backups

The use of Elasticsearch in GitLab is only ever as a secondary data store. This means that all of the data stored in Elasticsearch can always be derived again from other data sources, specifically PostgreSQL and Gitaly. Therefore if the Elasticsearch data store is ever corrupted for whatever reason you can reindex everything from scratch.

If your Elasticsearch index is incredibly large it may be too time consuming or cause too much downtime to reindex from scratch. There aren't any built in mechanisms for automatically finding discrepancies and resyncing an Elasticsearch index if it gets out of sync but one tool that may be useful is looking at the logs for all the updates that occurred in a time range you believe may have been missed. This information is very low level and only useful for operators that are familiar with the GitLab codebase. It is documented here in case it is useful for others. The relevant logs that could theoretically be used to figure out what needs to be replayed are:

All non-repository updates that were synced can be found in elasticsearch.log by searching for track_items and these can be replayed by sending these items again through ::Elastic::ProcessBookkeepingService.track!
All repository updates that occurred can be found in elasticsearch.log by searching for indexing_commit_range. Replaying these requires resetting the IndexStatus#last_commit/last_wiki_commit to the oldest from_sha in the logs and then triggering another index of the project using ElasticCommitIndexerWorker
All project deletes that occurred can be found in sidekiq.log by searching for ElasticDeleteProjectWorker. These updates can be replayed by triggering another ElasticDeleteProjectWorker.

With the above methods and taking regular Elasticsearch snapshots we should be able to recover from different kinds of data loss issues in a relatively short period of time compared to indexing everything from scratch.