Edit

Drop and rebuild an index in Azure AI Search

  • Article

This article explains how to drop and rebuild an Azure AI Search index. It explains the circumstances under which rebuilds are required, and provides recommendations for mitigating the impact of rebuilds on ongoing query requests. If you have to rebuild frequently, we recommend using index aliases to make it easier to swap which index your application is pointing to.

During active development, it's common to drop and rebuild indexes when you're iterating over index design. Most developers work with a small representative sample of their data to facilitate this process.

Modifications requiring a rebuild

The following table lists the modifications that require an index rebuild.

Action Description
Delete a field To physically remove all traces of a field, you have to rebuild the index. When an immediate rebuild isn't practical, you can modify application code to disable access to the "deleted" field or use the $select query parameter to choose which fields are represented in the result set. Physically, the field definition and contents remain in the index until the next rebuild, when you apply a schema that omits the field in question.
Change a field definition Revising a field name, data type, or specific index attributes (searchable, filterable, sortable, facetable) requires a full rebuild.
Assign an analyzer to a field Analyzers are defined in an index and then assigned to fields. You can add a new analyzer definition to an index at any time, but you can only assign an analyzer when the field is created. This is true for both the analyzer and indexAnalyzer properties. The searchAnalyzer property is an exception (you can assign this property to an existing field).
Update or delete an analyzer definition in an index You can't delete or change an existing analyzer configuration (analyzer, tokenizer, token filter, or char filter) in the index unless you rebuild the entire index.
Add a field to a suggester If a field already exists and you want to add it to a Suggesters construct, you must rebuild the index.
Switch tiers In-place upgrades aren't supported. If you require more capacity, you must create a new service and rebuild your indexes from scratch. To help automate this process, you can use the index-backup-restore sample code in this Azure AI Search .NET sample repo. This app will back up your index to a series of JSON files, and then recreate the index in a search service you specify.

Modifications with no rebuild requirement

Many other modifications can be made without impacting existing physical structures. Specifically, the following changes don't require an index rebuild. For these changes, you can update an existing index definition with your changes.

  • Add a new field
  • Set the retrievable attribute on an existing field
  • Update searchAnalyzer on a field having an existing indexAnalyzer
  • Add a new analyzer definition in an index (which can be applied to new fields)
  • Add, update, or delete scoring profiles
  • Add, update, or delete CORS settings
  • Add, update, or delete synonymMaps
  • Add, update, or delete semantic configurations

When you add a new field, existing indexed documents are given a null value for the new field. On a future data refresh, values from external source data replace the nulls added by Azure AI Search. For more information on updating index content, see Add, Update or Delete Documents.

How to rebuild an index

During development, the index schema changes frequently. You can plan for it by creating indexes that can be deleted, recreated, and reloaded quickly with a small representative data set.

For applications already in production, we recommend creating a new index that runs side by side an existing index to avoid query downtime. Your application code provides redirection to the new index.

  1. Determine whether a rebuild is required. If you're just adding fields, or changing some part of the index that is unrelated to fields, you might be able to simply update the definition without deleting, recreating, and fully reloading it.

  2. Get an index definition in case you need it for future reference.

  3. Drop the existing index, assuming you aren't running new and old indexes side by side.

    Any queries targeting that index are immediately dropped. Remember that deleting an index is irreversible, destroying physical storage for the fields collection and other constructs. Pause to think about the implications before dropping it.

  4. Create a revised index, where the body of the request includes changed or modified field definitions.

  5. Load the index with documents from an external source.

When you create the index, physical storage is allocated for each field in the index schema, with an inverted index created for each searchable field. Fields that aren't searchable can be used in filters or expressions, but don't have inverted indexes and aren't full-text or fuzzy searchable. On an index rebuild, these inverted indexes are deleted and recreated based on the index schema you provide.

When you load the index, each field's inverted index is populated with all of the unique, tokenized words from each document, with a map to corresponding document IDs. For example, when indexing a hotels data set, an inverted index created for a City field might contain terms for Seattle, Portland, and so forth. Documents that include Seattle or Portland in the City field would have their document ID listed alongside the term. On any Add, Update or Delete operation, the terms and document ID list are updated accordingly.

Balancing workloads

Indexing doesn't run in the background, but the search service will balance any indexing jobs against ongoing queries. During indexing, you can monitor query requests in the portal to ensure queries are completing in a timely manner.

If indexing workloads introduce unacceptable levels of query latency, conduct performance analysis and review these performance tips for potential mitigation.

Check for updates

You can begin querying an index as soon as the first document is loaded. If you know a document's ID, the Lookup Document REST API returns the specific document. For broader testing, you should wait until the index is fully loaded, and then use queries to verify the context you expect to see.

You can use Search Explorer or a Web testing tool like Postman to check for updated content.

If you added or renamed a field, use $select to return that field: search=*&$select=document-id,my-new-field,some-old-field&$count=true

See also