> The summarization during VACUUM invokes IndexBuildHeapRangeScan() which is
> passed scanStartBlock and scanNumBlocks. If scanStartBlock + scanNumBlocks
>> heapTotalBlocks, further down the line, heapgettup() may start returning
> tuples from the beginning given the following code in it:
>
> page++;
> if (page >= scan->rs_nblocks)
> page = 0;
>
> finished = (page == scan->rs_startblock) ||
> (scan->rs_numblocks != InvalidBlockNumber ?
> --scan->rs_numblocks == 0 :
> false);
>
> Where finished indicates whether it thinks the end of heap is reached.
>
> In this case, scan->rs_startblock is 384 set by IndexBuildHeapRangeScan()
> using heap_setscanlimits(). One can imagine how the above heap finish
> criteria might not work as expected.
What scares me is:
1) the bug will not be found unless someone inspects the internal data of BRIN. Regression test is useless here.
2) the bug effectively causes vacuum scans the heap *twice*, which will produce lots of I/O if the heap is not small.
3) I wonder if other index type is suffered by this type of bug.
I vaguely recall that more internal inspecting type brin regression
test was removed because it depends on contrib modules. I am not sure
whether the decision was appropreate or not.
Best regards,
--
Tatsuo Ishii
SRA OSS, Inc. Japan
English: http://www.sraoss.co.jp/index_en.php
Japanese:http://www.sraoss.co.jp