When executing SQL, you must refer to the definition of the table or index you want to access. These definitions are cached in the local memory of the backend process separately from the shared memory buffer of the database on shared_buffers because they are referenced each time SQL is executed. The direct definition is a tuple of system catalogs. The cache for this tuple is called "Catalog Cache". A structure that makes the definition easy to use is called a "Relations Cache". And in Fujitsu Enterprise Postgres, these two are collectively called "Meta Cache".

The meta cache will be kept indefinitely for performance reasons. In a large database, a single backend process accesses a large number of tables and so on, which results in a large meta-cache for the backend process. As a result, the sum of the local memory of the backend process may exceed the realistic memory size.

On the other hand, the feature to reduce the meta cache for the entire instance by sharing the meta cache between backend processes is the Global Meta Cache feature. The current Global Meta Cache feature only shares the catalog cache. Therefore, the metacache in Global Meta Cache now refers to the catalog cache.

What you still cannot share using the current Global Meta Cache feature and need to keep in local memory is the information (Meta cache header) and relation cache to access Global Meta Cache in shared memory. If you do not use the Global Meta Cache feature, keep the catalog and relation caches in local memory. The meta cache held in local memory is called the "Local Meta Cache". The feature to limit the size of a Local Meta Cache by removing it if it has not been accessed for a long time is the Local Meta Cache limit feature.

The Global Meta Cache feature and the Local Meta Cache limit feature can be used together to provide the strictest control over memory consumption. Of course, you can use only one or the other.

However, the Global Meta Cache feature has several percent overhead to access shared memory. The Local Meta Cache limit feature also causes the overhead of reholding the metacache because it may discard the previously held metacache. Therefore, consider using these features when your estimates do not allow for memory consumption.