Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Foundry IQ (Azure AI Search) is a fully managed search service that indexes content and supports retrieval through APIs, applications, and AI agents. It's suitable for enterprise search scenarios and AI-powered customer experiences that require dynamic content generation through chat completion models. As an Azure service, Foundry IQ provides a range of capabilities to support your reliability requirements.
When you use Azure, reliability is a shared responsibility. Microsoft provides a range of capabilities to support resiliency and recovery. You're responsible for understanding how those capabilities work within all of the services you use, and selecting the capabilities you need to meet your business objectives and uptime goals.
This article describes how Foundry IQ handles transient faults, availability zone failures, region-wide failures, and service maintenance. It also describes backup and restore options and the configuration requirements for the Foundry IQ service-level agreement (SLA).
Production deployment recommendations for reliability
For production workloads, we recommend that you:
Use a billable tier. Don't use the Free tier, which doesn't provide an SLA.
If you use the Dedicated pricing model, configure at least two replicas for read-only workloads, or at least three replicas for read-write workloads. Multiple replicas improve resilience to transient faults and service maintenance and meet the applicable SLA requirements.
Deploy your search service in a region that supports availability zones, so that Foundry IQ can distribute capacity across zones.
Reliability architecture overview
When you use Foundry IQ, you create a search service. Each search service supports many search indexes that store your searchable content.
Foundry IQ isn't designed as a primary data store. You populate search indexes by pushing documents from your application or by using indexers, which retrieve content from supported external data sources. Indexers can invoke skills to process and enrich content before adding it to an index.
Foundry IQ offers two pricing models, Dedicated and Serverless. Both provide the same core search features but manage capacity differently.
Dedicated pricing model
In the Dedicated pricing model, you select a service tier and configure the replicas and partitions for your service. You're responsible for provisioning enough capacity for your workload. Each replica holds a copy of the index data that the search engine runs retrieval against. Each replica represents a single virtual machine (VM). Basic services support up to three replicas. Standard and Storage Optimized services support up to 12 replicas. For more information, see Service limits in Foundry IQ.
The addition of multiple replicas allows Foundry IQ to:
Increase the availability of the search service.
Perform maintenance on one replica while queries continue to run on other replicas.
Handle higher throughput of indexing and query workloads.
Improve resiliency by attempting to provision replicas in different availability zones, if the region supports them.
Foundry IQ automatically assigns one replica to be the primary replica. All write operations are performed against that replica. The other replicas are used for read operations.
The following diagram illustrates how a search service with three replicas might be spread across three availability zones:
You can also configure the number of partitions, which represent the storage that the search indexes use.
It's important to understand the impact of adding replicas and partitions because they each affect read and write performance in different ways. For more information about replicas and partitions, see Estimate and manage capacity of a search service.
Serverless pricing model
When you use the Serverless pricing model, you don't need to configure replicas or partitions. Foundry IQ manages capacity for you based on query volume, indexing activity, and workload complexity. The reliability capabilities in this article, such as zone redundancy, apply to Serverless services too. Guidance about replica and partition counts applies only to Dedicated services.
Dependencies
Depending on the features that you use, your search service might require other resources and models. A failure in a dependency can affect indexing or queries even when your search service is healthy. You're responsible for configuring and managing the resources and endpoints that you provide. Microsoft manages the underlying infrastructure for Azure services and the models that Foundry IQ hosts.
Data sources: Indexers retrieve content from external data sources. You also need your source data to rebuild an index.
Skills and models that you configure: Skills can call external services, model deployments, or web APIs that you host to perform AI enrichment during indexing. Queries can also depend on models that you configure. For example, agentic retrieval, which combines results from multiple queries, can use a model for query planning and answer generation.
A vectorizer calls an embedding model that you configure to convert query input into vectors. If the vectorizer can't obtain a response from the model after retries, the entire query fails, including hybrid queries. Foundry IQ doesn't automatically fall back to keyword-only results.
Azure Storage: Some optional features write data to a storage account that you provide. An enrichment cache (preview) stores processed content for reuse between indexing runs. A knowledge store saves enriched content to tables or blobs for other applications to use. You're responsible for configuring the storage account's reliability capabilities. For more information, see Reliability in Azure Blob Storage and Reliability in Azure Table Storage.
Microsoft-hosted models: Semantic ranker reranks search results, and query rewrite (preview) generates alternative queries. Both features use models that Microsoft hosts and manages.
By default, semantic ranker returns partial results if its model is unavailable. You can configure requests to fail instead. If query rewrite is unavailable, Foundry IQ uses the original query text.
Resilience to transient faults
Transient faults are short, intermittent failures in components. They occur frequently in a distributed environment like the cloud, and they're a normal part of operations. Transient faults correct themselves after a short period of time. It's important that your applications can handle transient faults, usually by retrying affected requests.
All cloud-hosted applications should follow the Azure transient fault handling guidance when they communicate with any cloud-hosted APIs, databases, and other components. For more information, see Recommendations for handling transient faults.
Search services might experience transient faults during maintenance operations. For more information, see Resilience to service maintenance. Transient faults can also occur when the search service is indexing or when your application queries the index.
Foundry IQ indexers have built-in transient fault handling. If a data source is briefly unavailable, the indexer is designed to recover and retry. It uses change tracking to resume indexing from the last successfully indexed document.
If you build any applications that interact with Foundry IQ, they should handle transient faults. Use a retry strategy with exponential backoff for both read and write operations.
Resilience to availability zone failures
Availability zones are physically separate groups of datacenters within an Azure region. When one zone fails, services can fail over to one of the remaining zones.
When your search service meets the requirements, Foundry IQ is zone redundant; it distributes capacity across multiple availability zones within the search service region.
The availability zone behavior that this section describes applies to both Dedicated and Serverless services, but the configuration is different:
For Serverless services: Foundry IQ manages the replicas for you. Each Serverless service provides a read-write SLA because it's backed by the equivalent of three replicas.
For Dedicated services: When you add two or more replicas, Foundry IQ attempts to place each replica in a different availability zone. For services that have more replicas than available zones, replicas are distributed across zones as evenly as possible.
The following diagram illustrates how an example search service with four replicas might be deployed across three availability zones:
If your Dedicated service doesn't meet these requirements, it isn't zone redundant. For example, a Dedicated service with one replica might become unavailable if the availability zone that hosts that replica has an outage.
Important
Foundry IQ doesn't guarantee the exact zone placement of replicas. Placement is subject to capacity constraints, scaling operations, and other factors.
Requirements
Zone redundancy is automatically enabled when your search service meets all of the following criteria:
Region support: Support for availability zones depends on infrastructure and storage. For a list of supported regions, see Choose a region for Foundry IQ.
Tier: For the Dedicated pricing model, use the Basic tier or higher.
For the Serverless pricing model, Foundry IQ manages zone redundancy for you.
Number of replicas: For the Dedicated pricing model, use at least two replicas.
For the Serverless pricing model, Foundry IQ manages replicas for you.
Note
For Dedicated services, Foundry IQ attempts to distribute replicas across multiple zones when you have two or more replicas. However, you need at least three replicas to receive a read-write SLA. Serverless services don't use replica counts and every Serverless service has a read-write SLA for its lifetime. For more information, see Service-level agreement.
Instance distribution across zones
For the Dedicated pricing model, Foundry IQ attempts to place replicas across different availability zones. However, there are occasionally situations where all of the replicas of a search service might be placed into the same availability zone. This situation can happen when replicas are removed from your service, such as when you scale in by configuring your service to use fewer replicas. Replica removal doesn't trigger the remaining replicas to rebalance across the availability zones.
To reduce the likelihood of all of your replicas being placed into a single availability zone, you can manually trigger a scale-out operation immediately after a scale-in operation. For example, suppose that your search service has 10 replicas and you want to scale in to 7 replicas. Instead of performing a single scale operation, you can temporarily scale to 6 replicas and then immediately scale to 7 replicas to trigger zone rebalancing.
Cost
Each Dedicated search service starts with one replica. Zone redundancy requires two or more replicas, which increases the cost to run the service. To understand the billing implications of replicas, use the pricing calculator.
With the Serverless pricing model, you don't need to scale your service to get zone redundancy. Zone redundancy and the SLA are included. You pay consumption-based charges for compute and storage.
Configure availability zone support
If your search service meets the requirements for zone redundancy, no extra configuration is necessary. Whenever possible, Foundry IQ attempts to place your replicas in different availability zones.
Capacity planning and management
If you use the Dedicated pricing model, prepare for availability zone failure by considering overprovisioning the number of replicas. Overprovisioning allows the search service to tolerate some capacity loss and continue to function without degraded performance. Adding replicas during an outage is challenging, so overprovisioning helps ensure that your search service can handle normal request volumes, even with reduced capacity. For more information, see Manage capacity by overprovisioning.
With the Serverless pricing model, you don't need to plan for overprovisioning. Serverless billing is based on your usage. Foundry IQ manages capacity on the backend to meet your needs.
Behavior when all zones are healthy
This section describes what to expect when search services are configured for zone redundancy and all availability zones are operational.
Cross-zone operation: Foundry IQ automatically balances queries across the available replicas. It can send read operations to any replica in any availability zone. It sends write operations to a single primary replica that the Foundry IQ service selects.
Cross-zone data replication: Changes in data are automatically replicated between replicas across availability zones. Replication occurs asynchronously, which means that writes are committed to one primary replica before they're replicated to other replicas.
Behavior during a zone failure
This section describes what to expect when search services are configured for zone redundancy and an availability zone outage occurs.
- Detection and response: Foundry IQ is responsible for detecting a failure in an availability zone. You don't need to do anything to initiate a zone failover.
- Notification: Microsoft doesn't automatically notify you when a zone is down. However, you can use Azure Resource Health to monitor the health of an individual resource, and you can set up Resource Health alerts to notify you of problems. You can also use Azure Service Health to understand the overall health of the service, including any zone failures, and you can set up Service Health alerts to notify you of problems.
Active requests: Requests that replicas process in the failed zone are terminated. Clients should retry the requests by following the guidance for handling transient faults.
Expected data loss: If the affected availability zone only contains read replicas, no data loss is expected.
If the primary replica is lost because it was in the affected zone, then any write operations that haven't yet been replicated might be lost.
Expected downtime: In most situations, a zone failure isn't expected to cause downtime to your search service for read operations because read replicas in other availability zones continue to serve requests.
If the primary replica is lost because it was in the affected zone, Foundry IQ automatically promotes another replica to become the new primary so that write operations can resume. It typically takes a few seconds for the replica promotion to occur. During this time, write operations might not succeed. Ensure that your applications are prepared by following transient fault handling guidance.
However, there are some unlikely situations where all of your search service's replicas might be in a single availability zone. In this scenario, you might experience downtime until the zone recovers. For more information, and to understand a workaround, see Instance distribution.
Redistribution: When a zone fails, Foundry IQ detects the failure and routes requests to active replicas in the surviving zones. If the primary replica is lost, another replica is promoted to be the new primary.
Zone recovery
When the availability zone recovers, Foundry IQ automatically restores normal operations and begins routing traffic to available replicas across the zones used by the service, including the recovered zone.
Test for zone failures
Foundry IQ manages traffic routing for zone-redundant services. You don't need to initiate or validate any zone failure processes.
Resilience to region-wide failures
Foundry IQ is a single-region service. If the region becomes unavailable, your search service also becomes unavailable.
Custom multiregion solutions for resiliency
You can deploy multiple Foundry IQ services in different regions to improve resilience to region-wide failures. Foundry IQ doesn't automatically replicate indexes or fail over between regions.
You're responsible for deploying and configuring the services, synchronizing their indexes, and configuring request routing and failover policies. Plan capacity in the secondary region before an outage. Regional capacity constraints can prevent you from creating a service when you need it. For more information, see Plan Foundry IQ deployments in high-demand regions.
To optimize the performance of your overall solution, look for opportunities to perform indexing on read-only replicas of your data sources. For example, some indexers support reading from a geo-distributed data source's read replicas.
For more information, see Multi-region deployments in Foundry IQ.
Backup and restore
Because Foundry IQ isn't a primary data storage solution, it doesn't provide self-service backup and restore options. However, you can use the index-backup-restore sample for .NET or Python to back up your index definition and its documents to a series of JSON files, which are then used to restore the index. Review each sample's prerequisites and limitations, and test restoring an exported index before relying on it for recovery.
Deleting an index is irreversible. If you accidentally delete an index and don't have a backup, you can rebuild the index. Rebuilding involves recreating the index on your search service and then reloading it by retrieving data from your primary data store. You need the original source data to rebuild the index.
To help prevent accidental deletion of your search service, you can apply a Delete resource lock. The lock protects the search service resource, but it doesn't protect individual indexes or indexed documents from deletion.
Resilience to service maintenance
Search services might experience transient faults during standard, unscheduled maintenance operations. Foundry IQ doesn't provide advance notification or allow scheduling of maintenance at specific times. Although every effort is made to minimize downtime, even for single-replica services, brief interruptions can still occur.
Your applications should also follow the guidance for handling transient faults.
For the Dedicated pricing model, some operations that you initiate also affect your service:
Service upgrades: With two or more replicas, your service remains available during a service upgrade while one replica is updated.
Scaling and pricing tier changes: Supported changes to replica counts, partition counts, and pricing tiers run in the background. Your service remains available for read and write operations. These operations can take several hours to complete, and you can't cancel them. For more information, see Manage and adjust capacity.
Service-level agreement
The service-level agreement (SLA) for Azure services describes the expected availability of each service and the conditions that your solution must meet to achieve that availability expectation. For more information, see SLAs for online services.
Foundry IQ doesn't provide an availability SLA for the Free tier.
For billable services, the SLA requirements depend on the pricing model:
- Dedicated pricing model: Configure at least two replicas for a read SLA (queries), or at least three replicas for a read-write SLA (queries and indexing).
- Serverless pricing model: Every service has a read-write SLA for its lifetime. You don't need to configure replicas.