Purpose of OpenTofu State

State is a necessary requirement for OpenTofu to function. It is often asked if it is possible for OpenTofu to work without state, or for OpenTofu to not use state and just inspect real world resources on every run. This page will help explain why OpenTofu state is required.

As you'll see from the reasons below, state is required. And in the scenarios where OpenTofu may be able to get away without state, doing so would require shifting massive amounts of complexity from one place (state) to another place (the replacement concept).

Mapping to the Real World​

OpenTofu requires some sort of database to map OpenTofu config to the real world. For example, when you have a resource resource "aws_instance" "foo" in your configuration, OpenTofu uses this mapping to know that the resource resource "aws_instance" "foo" represents a real world object with the instance ID i-abcd1234 on a remote system.

For some providers like AWS, OpenTofu could theoretically use something like AWS tags. Early prototypes of OpenTofu actually had no state files and used this method. However, we quickly ran into problems. The first major issue was a simple one: not all resources support tags, and not all cloud providers support tags.

Therefore, for mapping configuration to resources in the real world, OpenTofu uses its own state structure.

OpenTofu expects that each remote object is bound to only one resource instance in the configuration. If a remote object is bound to multiple resource instances, the mapping from configuration to the remote object in the state becomes ambiguous, and OpenTofu may behave unexpectedly. OpenTofu can guarantee a one-to-one mapping when it creates objects and records their identities in the state. When importing objects created outside of OpenTofu, you must make sure that each distinct object is imported to only one resource instance.

Metadata​

Alongside the mappings between resources and remote objects, OpenTofu must also track metadata such as resource dependencies.

OpenTofu typically uses the configuration to determine dependency order. However, when you delete a resource from an OpenTofu configuration, OpenTofu must know how to delete that resource from the remote system. OpenTofu can see that a mapping exists in the state file for a resource not in your configuration and plan to destroy. However, since the configuration no longer exists, the order cannot be determined from the configuration alone.

To ensure correct operation, OpenTofu retains a copy of the most recent set of dependencies within the state. Now OpenTofu can still determine the correct order for destruction from the state when you delete one or more items from the configuration.

One way to avoid this would be for OpenTofu to know a required ordering between resource types. For example, OpenTofu could know that servers must be deleted before the subnets they are a part of. The complexity for this approach quickly explodes, however: in addition to OpenTofu having to understand the ordering semantics of every resource for every provider, OpenTofu must also understand the ordering across providers.

OpenTofu also stores other metadata for similar reasons, such as a pointer to the provider configuration that was most recently used with the resource in situations where multiple aliased providers are present.

Performance​

In addition to basic mapping, OpenTofu stores a cache of the attribute values for all resources in the state. This is the most optional feature of OpenTofu state and is done only as a performance improvement.

When running a tofu plan, OpenTofu must know the current state of resources in order to effectively determine the changes that it needs to make to reach your desired configuration.

For small infrastructures, OpenTofu can query your providers and sync the latest attributes from all your resources. This is the default behavior of OpenTofu: for every plan and apply, OpenTofu will sync all resources in your state.

For larger infrastructures, querying every resource is too slow. Many cloud providers do not provide APIs to query multiple resources at once, and the round trip time for each resource is hundreds of milliseconds. On top of this, cloud providers almost always have API rate limiting so OpenTofu can only request a certain number of resources in a period of time. Larger users of OpenTofu make heavy use of the -refresh=false flag as well as the -target/-exclude flags in order to work around this. In these scenarios, the cached state is treated as the record of truth.

Syncing​

In the default configuration, OpenTofu stores the state in a file in the current working directory where OpenTofu was run. This is okay for getting started, but when using OpenTofu in a team it is important for everyone to be working with the same state so that operations will be applied to the same remote objects.

Remote state is the recommended solution to this problem. With a fully-featured state backend, OpenTofu can use remote locking as a measure to avoid two or more different users accidentally running OpenTofu at the same time, and thus ensure that each OpenTofu run begins with the most recent updated state.