The metadata for mini-blocks is very small (typically less than 64 bytes per 8MiB disk page). At initialization time we often have to issue many small reads to read the metadata. In practice this isn't quite as bad as it seems since we default to 1Mi rows per fragment and so most columns will never have more than one disk page. Still, we do have to perform one read per column. It might be worth it to either store the metadata in the fragment metadata or in some shared global buffer to reduce the number of IOPS that must be performed to initialize the search cache.
The metadata for mini-blocks is very small (typically less than 64 bytes per 8MiB disk page). At initialization time we often have to issue many small reads to read the metadata. In practice this isn't quite as bad as it seems since we default to 1Mi rows per fragment and so most columns will never have more than one disk page. Still, we do have to perform one read per column. It might be worth it to either store the metadata in the fragment metadata or in some shared global buffer to reduce the number of IOPS that must be performed to initialize the search cache.