Dan Jackson explains the concept of latching in SQL Server:
To start with, a basic definition: ‘Latches are lightweight synchronization objects, that are used by the storage engine of SQL Server to protect the internal memory structures’. Compare this with locks in SQL server, which are a transaction level construct to manage concurrency, latches work at the thread level to maintain data integrity within the internal memory structures. They are not exposed outside of the SQL Server Operating System (SQLOS). They are only managed by SQL Server itself, not by users (unlike locks that can be overridden via lock hints or changing isolation level). It is useful to keep in mind that a single transaction can use multiple threads at the same time.
Latching is a funny concept to me, in that I think people say “This must be a latching problem” far more than there actually is, but when there is a proper latching problem, it usually winds up being a pretty big deal.