Bruce Momjian <pgman@candle.pha.pa.us> writes:
>> Basically, move the first 100 rows to the end of the table file, then take
>> 100 and write it to position 0, 101 to position 1, etc ... that way, at
>> max, you are using ( tuple * 100 ) bytes of disk space, vs 2x the table
>> size ... either method is going to lock the file for a period of time, but
>> one is much more friendly as far as disk space is concerned *plus*, if RAM
>> is available for this, it might even be something that the backend could
>> use up to -S blocks of RAM to do it off disk? If I set -S to 64meg, and
>> the table is 24Meg in size, it could do it all in memory?
> Yes, I liked that too.
What happens if you crash partway through?
I don't think it's possible to build a crash-robust rewriting ALTER
process that doesn't use 2X disk space: you must have all the old tuples
AND all the new tuples down on disk simultaneously just before you
commit. The only way around 2X disk space is to adopt some logical
renumbering approach to the columns, so that you can pretend the dropped
column isn't there anymore when it really still is.
regards, tom lane