Overall query volume on an Azure AI Search service is low, but one complex query that…

How it works

Azure AI Search scales in two directions, and they solve different problems:

• Replicas add capacity to the service as a whole.
• Partitions split the data across extra computing resources. Two partitions split it in half, three into thirds, and so on.

Microsoft's performance guidance addresses this exact case: if the problem is a single query that takes too long, adding replicas won't help, but more partitions might. A side effect of adding partitions is that slower queries sometimes run faster due to parallel computing. Microsoft has seen that parallelization on low-selectivity queries, such as queries that match many documents, or facets providing counts over a large number of documents. That's this query.

Facets have a cost of their own. Every faceted query needs aggregations, so only add facets you plan to render, and avoid requesting high facet counts unless you need them.

The other options, and when you'd use them

• High $skip paging: A high $skip makes things worse. The engine retrieves and ranks more documents per request, and Microsoft recommends filtering instead.
• Rewrite as a regular expression: Regex queries can be particularly expensive and need full index scans. They add work rather than remove it.
• Add replicas: Replicas add overall capacity. Total volume here is low, and for one long-running query adding replicas won't help.

Exam tip

One slow, heavy query = partitions (parallelism). Many queries slow under load = replicas.

References

• Tips for better performance in Azure AI Search