The block is the structure that holds row data. The Logical Storage Structure documentation has a detailed overview of all this.
Note that data for a single row can reside on more than one block (this is called row chaining), so it need not be contiguous on disk. Row migration can also leave just a "forwarding address" in the row's original block that points to the row's new block of residence.
All I/O to and from the SGA is done in block-sized chunks (possibly more than one block at a time, but, for the SGA, never smaller than a block).
The block size of the SYSTEM tablespace determines the database's "natural" block size. You can have other tablespaces with different block sizes but:
- you can't mix block sizes inside a same tablespace
- you will need to set up separate cache areas for each block size (you can't have for example 8k blocks and 16k blocks in the same block cache)
So in general, it is simpler to use only one block size, if only to avoid the administrative and tuning overhead of managing the size of multiple block buffers.
How you determine the optimal block size for a database is a compromise just like pretty much everything else.
Smaller blocks means your SGA is more granular, but each block has relatively more overhead. And wide rows in small blocks will be more likely to chain or migrate which can be a source of performance problems.
Larger blocks means less chances of chaining (assuming PCTFREE and PCTUSED are reasonably set). But it also means large single-block I/Os and less granularity in your SGA. Since a change in any row of one block requires the whole block to be written out to disk, larger block sizes could lead to more write I/O volume than smaller ones.
Another factor to take into account is that traditional (not bigfile) datafiles are limited to 4M blocks. That's about 32G for 8k blocks, but only 8G for 2k blocks. Since the number of files per (non-bigfile) tablespace is limited (to 1022), this could be important. (There is also a maximum of 65533 files per database, possibly less. See physical database limits.)
Transactional systems tend to use relatively smaller block sizes than datawarehousing systems, but that really isn't anything like a hard and fast rule. (e.g. current Oracle best practice guidelines on Exadata is 8k blocks regardless of scenarios.).