It depends. We have three basic ways to accomplish reading this table. 1) direct path read, 2) conventional path read using the buffer cache - but only using a small portion of it (since we cannot cache the entire table anyway) and 3) by a conventional path read into the buffer cache but attempting to cache everything we can (by getting rid of the other stuff in the cache)


We could choose the read the table into your PGA bypassing the SGA altogether. This is called a direct path read.

We could choose to read the table via the buffer cache. By default - a large table such as this will be loaded only into a small portion of the buffer cache and aged out quickly so as to not completely flush the buffer cache. So, we'll read a bit of the table, put it into the cache, access and process it from the cache - read another bit and do the same - and flush out some of the earlier bits we read to make room for more (instead of flushing out all of the other stuff we have cached)

You could "alter table t cache" (but probably shouldn't!!!) - and that would cause us to read as much of the table as possible into the buffer cache - flushing out the other blocks that were there. that would be unproductive in this case since only the last 20gb would be in the cache and the next full scan would flush out that 20gb to get the first 20gb of the table into the cache (and then flush that out to get the next 20 and so on)


In the most usual case - the table would be paged into the SGA a small bit at a time and aged out rapidly to make room for another bit of that table. In that fashion - we don't wipe all of the other good stuff out of the cache.