> ## Documentation Index
> Fetch the complete documentation index at: https://docs.push.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Caching

> Understand how caching and datasource queries work in Push.ai.

Push.ai uses a caching strategy to optimize application performance, while reducing the load on your data sources.

## How it works?

Push.ai queries datasources through two methods:

1. **Calculations**: metric and dimension queries using one of our [calculations](/business-objects/metrics#calculations).
2. **Time-series**: metric and dimension queries using a daily time grain.

## When is caching used?

Caching is used when querying both calculations and timeseries data. Push.ai implements a caching strategy for each unique metric-dimension combination. For each combination, we maintain a daily cache, meaning that we will query that specific combination from the data source at most once per day.

## When are datasource queries triggered?

Push.ai queries datasources using two methods:

1. **In-app Exploration**: Datasource queries are triggered when you interact with a metrics and dimension in the Push.ai UI.
2. **Scheduled Updates**: Datasource queries are triggered on a schedule defined in both Reviews and Subscriptions. With Subscriptions, datasource queries are triggered for each Business Object a user is subscribed to, in addition to any related Business Objects. Related objects include any related metrics and are limited to [AI-enabled dimensions](/business-objects/dimensions#ai-enabled-dimensions) only.

<Info>
  If multiple users are subscribed to the same Business Object, the datasource
  query will be triggered once for each object, rather than once for each user.
</Info>

## When is datasource compute not used?

Datasource compute is not used in any of Push.ai's advanced analytics or AI applications. This includes time-series modeling, such as outlier detection and forecasting. In addition, Push.ai's AI systems combine the existing calculations with large language models, and do not use additional datasource compute.