Reindex an Elasticsearch index
After updating an Elasticsearch index’s schema by changing the fields or document types, you need to reindex the affected index before the new fields and types can be used.
The reindexing process:
- Locks the Elasticsearch index to prevent writes to the index while data is being copied
- Creates a new index using the schema defined in the deployed version of search-api
- Copies all the data from the old to the new index
- Compares the old and new data to check for inconsistencies
- If everything looks the same, switches the alias to the new index
Note: If you have changed the govuk_document_types gem, you do not need
to migrate the schema for your changes to be applied to documents in
elasticsearch. Instead, you can run the rake task search:update_supertypes
to update documents inplace without locks. This can be done during working hours.
How to reindex an Elasticsearch index
Do not reindex on production during working hours except in an emergency. Reindexing locks the index for writes, so content is not updated in the search index. See the Replay traffic section below if you need to run a reindexing during working hours. Reindexing takes around 2 hours to complete.
To reindex, run the search:migrate_schema rake task:
bundle exec rake search:migrate_schema SEARCH_INDEX=alias_of_index_to_migrate
If you set the last parameter to SEARCH_INDEX=all, search-api will reindex all
the indices sequentially.
You can run this task from Jenkins, but it will block other rake tasks from
being run for 15 minutes to an hour. You can avoid this by running the command
directly on a search machine, but you need to prefix the command with
govuk_setenv search-api to make sure the Elasticsearch hostname is set
correctly.
To monitor progress, SSH to a search box:
ssh $(ssh integration "govuk_node_list --single-node -c search").integration
Then check how many documents have been copied to the new index:
curl http://elasticsearch6/_cat/indices?v
Replay traffic
This step is only necessary if you ran reindexing job during working hours, which means that content published in whitehall will be missing from search.
See Replaying traffic to correct an out of sync search index for details.
Cleanup
Reindexing does not delete the old index. This lets us switch back to the old index if there is a serious problem with the new one.
Once you’re confident that the reindexing was successful, delete the old (unaliased) index using the search-api rake task:
rake search:clean SEARCH_INDEX=alias_of_index_to_clean_up
Avoid leaving old indices around for more than a few days. Performance starts to degrade once there are more than three or four old indices in the cluster, and if enough old indices hang around, we may hit space limitations and be unable to index new documents.
Troubleshooting
To stop the reindexing job
If you need to cancel the reindexing while it’s in progress:
- Stop the reindexing rake task
- Unlock the old index by running the search-api rake task:
rake search:unlock SEARCH_INDEX=alias_of_index_to_unlock
This doesn’t actually stop the reindexing, because reindexing is an internal Elasticsearch progress triggered by the rake task. It will stop the rake task from switching the alias over to the new index once it has copied all the data, which is normally good enough.
If you need to stop the reindexing process itself, for example because Elasticsearch is about to run out of disk space, connect to the search box (see above) then:
Find the ID of the reindexing task:
curl 'http://elasticsearch6/_tasks?actions=%2Areindex&pretty'Stop the task:
curl -XPOST http://elasticsearch6/_tasks/{task_id}/_cancel
To switch back to the old index
If you discover a problem after reindexing and need to switch back to the old index, run this search-api rake task:
rake search:switch_to_named_index[full_index_name] SEARCH_INDEX=index_alias
where full_index_name is the full name of the new index, including the date
and UUID, e.g. govuk-2018-01-29t17:08:21z-31f39bdb-c62b-4607-8081-19ea87fb1498.
Switching back to an old index means that you’ll lose any content updates that were published while the new index was live. To fix this, replay the traffic from both publishing-api and Whitehall.