"Tom Lane" <tgl@sss.pgh.pa.us> wrote
> "Qingqing Zhou" <zhouqq@cs.toronto.edu> writes:
>> BEGIN TRANSACTION WITHOUT XLOG; /* forbidden vacuum, PITR etc */
>
>> So during this peroid, if any transaction failed, the only consequence is
>> add invisible garbage data.
>
> No, the likely consequence is irretrievable corruption of any table or
> index page touched by the transaction.
>
I guess I know (at least part) of what you mean. This is because we rely on
replay all the xlog no matter it belongs to a committed transaction or not.
Why? Because a failed transaction is not totally useless since later
transaction may reply on some physical thing it creates - for example, a new
page and its links of a btree. So for heap, there is(95% sure) no such
problem. Our heap redo algorithm can automatically add empty pages to a
heap. For index, there are problems, but I suspect they are solvable by not
bypassing these records ... if this is not totally-nonstarter, I will
investigate details of how to do it.
Regards,
Qingqing