Re: Finding out the aging tables for vacuuming to prevent data loss on transaction id wraparound

Subramanian Krishnan wrote:
> We are running PostgreSQL 9.4.14 and recently we received an alert for transaction id
> crossing the 1 billion mark. Since we can hit the transaction id wraparound issue for aging tables,
> we investigated by running the following queries:
>
> mpsdb=> select * from txid_current();
> txid_current
> --------------
> 1100826671
> (1 row)
>
> mpsdb=> SELECT datname, age(datfrozenxid) FROM pg_database ORDER BY 2 DESC LIMIT 20;
> datname | age
> -----------+------------
> mpsdb | 1087909867
> rdsadmin | 1000562517
> postgres | 950697143
> template0 | 900709601
> template1 | 900697643
> (5 rows)
>
> mpsdb=> SELECT c.relname as table_name, c.relkind as type, age(c.relfrozenxid) as age, c.relfrozenxid FROM pg_class AS c WHERE age(c.relfrozenxid) <> 2147483647 ORDER BY 3 DESC LIMIT 20;
> table_name | type | age | relfrozenxid
> -------------------+------+------------+--------------
> pg_toast_22985917 | t | 1042877986 | 57941433
> pg_toast_22986301 | t | 1042877877 | 57941542
> pg_toast_23823161 | t | 1041791836 | 59027583
> pg_toast_23823242 | t | 1041791776 | 59027643
> pg_toast_23904594 | t | 1041658236 | 59161183
> pg_toast_23904629 | t | 1041658221 | 59161198
> pg_toast_23904640 | t | 1041658208 | 59161211
> pg_toast_23904648 | t | 1041658200 | 59161219
> pg_toast_23904658 | t | 1041658191 | 59161228
> pg_toast_23904666 | t | 1041658183 | 59161236
> pg_toast_23904674 | t | 1041658175 | 59161244
> pg_toast_23904682 | t | 1041658169 | 59161250
> pg_toast_23904690 | t | 1041658161 | 59161258
> pg_toast_23904698 | t | 1041658156 | 59161263
> pg_toast_23904706 | t | 1041658149 | 59161270
> pg_toast_23904714 | t | 1041658129 | 59161290
> pg_toast_23904722 | t | 1041658096 | 59161323
> pg_toast_23904730 | t | 1041658090 | 59161329
> pg_toast_23904738 | t | 1041658085 | 59161334
> pg_toast_23904746 | t | 1041658077 | 59161342
> (20 rows)
>
>
> mpsdb=> SELECT datname, usename, pid, waiting, current_timestamp - xact_start AS xact_runtime, query FROM pg_stat_activity WHERE upper(query) like '%VACUUM%' ORDER BY xact_start;
> datname | usename | pid | waiting | xact_runtime | query
> ---------+----------+-------+---------+-----------------+------------------------------------------------------------------------
> mpsdb | rdsadmin | 13833 | f | 00:00:14.161262 | autovacuum: VACUUM pg_toast.pg_toast_242649675 (to prevent wraparound)
> mpsdb | rdsadmin | 14565 | f | 00:00:11.735571 | autovacuum: VACUUM pg_toast.pg_toast_242649757 (to prevent wraparound)
> mpsdb | rdsadmin | 15315 | f | 00:00:09.303762 | autovacuum: VACUUM pg_toast.pg_toast_242649819 (to prevent wraparound)
> mpsdb | rdsadmin | 14935 | f | 00:00:06.893078 | autovacuum: VACUUM pg_toast.pg_toast_243226657 (to prevent wraparound)
> mpsdb | rdsadmin | 15851 | f | 00:00:04.322474 | autovacuum: VACUUM pg_toast.pg_toast_243227582 (to prevent wraparound)
> mpsdb | rdsadmin | 15615 | f | 00:00:01.768495 | autovacuum: VACUUM pg_toast.pg_toast_243226332 (to prevent wraparound)
>
> It is clear that we need to vacuum the most aged pg_toast tables which are not yet picked up by the autovacuum process.
>
> The question which is bothering us though is:
> As per the queries we ran the most aged transaction and table is 1042877986 (pg_toast_22985917).
> And as per PostgreSQL pg_database documentation the most aged table/transaction in pg_class determines
> age of the database (via the datfrozenxid value). If that is the case we would have expected the age
> of mpsdb database to be 1042877986 (or in this range) but what we see as age is 1087909867.
>
> Does this mean there are other aging tables we are not aware of? If yes, how do we get to those
> tables since we have already queried pg_class for the most aged table?
>
> Request any help/guidance we can get regarding this.

That is indeed strange.

Reading the code, I see that at the end of a VACUUM, "datfrozenxid" is set
to the minimum value of all "relfrozenxid" for all tables in the database
obtained with a sequential scan of "pg_class".

The only exception is that if a table is found where "relfrozenxid" is in the
future, nothing is done. The comment suggests that that should normally not
happen, but "has been known to arise due to bugs in pg_upgrade".

- Can you find tables with negative "relfrozenxid" age in "pg_class"?

- The other option is that a VACUUM finished between the query of "pg_database"
and the query of "pg_class". Is that an option?

You should tune autovacuum to be more aggressive so that it gets done processing
the tables. To get rid of the immediate problem, you could schedule a manual
VACUUM of the tables to which the TOAST table belong.

Yours,
Laurenz Albe