In Oracle, the Undo records allow data that is being modified to be accessed.
Oracle uses row level locking to prevent two processes from trying to update the same data at the same time.
Only the rows being modified are locked.
This means that the rest of the table can still be modified while a row is being changed. It also means that the whole of the table including the data that is currently being modified can still be read.
This is covered elsewhere but the key is that the data will always be consistent with its state at the time that an operation began.
The Users server process is the process that handles the users requests to the oracle database.
When the user requests a add, update or delete of data in the database, the server process modifies the data from the buffer cache, not the actual data files.
The DBWR (Database Writer) process is the only process that writes to the data files.
Here is the process by which data is modified in the Oracle database:
If the operation is an Insert:
Space found in a block in buffer cache.
Data inserted into that block buffer.
If the operation is an Update or a Delete:
The data is read, either from the buffer cache or the data files and if from the datafiles, it is placed in the buffer.
The data is deleted or modified in the buffer cache.
A copy of the origional data written to the UNDO tablespace in case it is needed for a rollback or for read consistency.
Once the data is modified in the Buffer Cache, a log file is written to the Redo log.
Once the Redo log is written out, the change is considered complete.
Later…the DBWR writes the changed blocks to the datafiles.
The DBWR is the only process that can write to datafiles but the Server Processes can read form them.
Reading from the Oracle RDMS is done by the users server process
Once a query is submitted, the users server process first determines if data is in the buffer cache of the SGA (System Global Area) memory area.
If the data is not already in the buffer cache, Oracle reads the data from the data files on the disk and writes it to the SGA.
The data is returned to the user or manipulated in memory from there, not from disk.
This means that commonly used data can be accessed form the Buffer Cache which is in RAM and very fast and not form the disk which is slow.
This is a very scalable architecture.