Thread: data file corruption
Hi All,
We are facing one strange problem about data file corruptions.
We have many postgres databases. At some point, one simple query on one database started crashing back-end.
The query is
select count(*), col1 from tab1 group by col1;
After using pg_filedump (http://pgfoundry.org/projects/pgfiledump/) on data files for tab1 (relnodeid in pg_class), we found that the number of attributes per tuple is different for few tuples in data file than the rest.
pg_filedump utility prints the contents of each tuple along with block header, data header.
In our case, the same data file has the following two data headers.
valid header
============
<Data> ------
Item 1 – Length: 114 Offset: 32648 (0x7f88) Flags: NORMAL
XMIN: 8849668 XMAX: 0 CID|XVAC: 0
Block Id: 0 linp Index: 1 Attributes: 10 Size: 24
infomask: 0x0902 (HASVARWIDTH|XMIN_COMMITTED|XMAX_INVALID)
Invalid header
==========
<Data> ------
Item 1 – Length: 234 Offset: 32528 (0x7f10) Flags: NORMAL
XMIN: 2959623 XMAX: 0 CID|XVAC: 0
Block Id: 155 linp Index: 1 Attributes: 92 Size: 40
infomask: 0x0903 (HASNULL|HASVARWIDTH|XMIN_COMMITTED|XMAX_INVALID)
As you can see, # of attributes and XMIN are different.
I have few questions regarding this.
1. When such case can occur?
2. Later we found that invalid header is actually valid header for other table's data file. So somehow data of one table got inserted into another and both tables have different # of attributes. Can this be possible? Any hardware issue can cause this?
3. Has anybody seen this problem?
Thanking you.
- Nachiket
PG User <pguser1982@gmail.com> writes: > ... Later we found that invalid header is actually valid header for other > table's data file. So somehow data of one table got inserted into another > and both tables have different # of attributes. Can this be possible? Any > hardware issue can cause this? There are (at least) 3 possibilities: 1. Postgres got confused and wrote a block to the wrong place. 2. The kernel (filesystem) got confused and wrote a block to the wrong place. 3. The disk hardware got confused and wrote a block to the wrong place. The easiest way to narrow it down would be to try to identify the size of the misplaced write. If postgres messed up, it'd certainly have written a whole 8K page to the wrong place. A filesystem bug would more likely have misplaced a single filesystem block, which might be 8K but I think it's more usually 1K or 4K on modern machines (a little research about your OS should tell you what blocksize is being used on this filesystem). Or if the disk screwed up, it'd most likely have misplaced a single 512-byte sector. So take a closer look at the pg_filedump results for the questionable page, and see if you can determine how much of the page looks to have been interpolated from someplace else. FWIW, if I had to bet with no further data, I'd bet on door #2. Bugs of this sort have been found in Postgres, but not in a long time. Is your kernel up-to-date? regards, tom lane
PG User <pguser1982@gmail.com> writes:
> ... Later we found that invalid header is actually valid header for other> table's data file. So somehow data of one table got inserted into anotherThere are (at least) 3 possibilities:
> and both tables have different # of attributes. Can this be possible? Any
> hardware issue can cause this?
1. Postgres got confused and wrote a block to the wrong place.
2. The kernel (filesystem) got confused and wrote a block to the wrong
place.
3. The disk hardware got confused and wrote a block to the wrong place.
The easiest way to narrow it down would be to try to identify the size
of the misplaced write. If postgres messed up, it'd certainly have
written a whole 8K page to the wrong place. A filesystem bug would more
likely have misplaced a single filesystem block, which might be 8K but
I think it's more usually 1K or 4K on modern machines (a little research
about your OS should tell you what blocksize is being used on this
filesystem). Or if the disk screwed up, it'd most likely have misplaced
a single 512-byte sector. So take a closer look at the pg_filedump
results for the questionable page, and see if you can determine how much
of the page looks to have been interpolated from someplace else.
FWIW, if I had to bet with no further data, I'd bet on door #2.
Bugs of this sort have been found in Postgres, but not in a long time.
Is your kernel up-to-date?
regards, tom lane