auto-vacuum & Negative "anl" Values

Greetings all,

I have been spending some time looking into how auto-vacuum is
performing on one of our servers. After putting the PostgreSQL logs
in debug I noticed that the threshold for ANALYZE was never being hit
for a particular table because the calculated value becomes
increasingly negative.

We have an entry in the pg_autovacuum table for a table that has
quite frequent inserts and updates, but not deletes. Here are the
values:

-[ RECORD 1 ]----+------
vacrelid | #####
enabled | t
vac_base_thresh | 500
vac_scale_factor | 0.1
anl_base_thresh | 200
anl_scale_factor | 0.05
vac_cost_delay | -1
vac_cost_limit | -1

I've noticed that the threshold for ANALYZE never gets met because
the threshold for VACUUM is hit first, therefore resetting the
counters. Here is a snippet of the log that shows what's happening:

DEBUG: tablename: vac: 961 (threshold 14217), anl: -9756 (threshold
7058)
DEBUG: tablename: vac: 1924 (threshold 14217), anl: -8792 (threshold
7058)
DEBUG: tablename: vac: 2953 (threshold 14217), anl: -7763 (threshold
7058)
DEBUG: tablename: vac: 3998 (threshold 14217), anl: -6718 (threshold
7058)
DEBUG: tablename: vac: 5170 (threshold 14217), anl: -5546 (threshold
7058)
DEBUG: tablename: vac: 6405 (threshold 14217), anl: -4311 (threshold
7058)
DEBUG: tablename: vac: 7635 (threshold 14217), anl: -3081 (threshold
7058)
DEBUG: tablename: vac: 8818 (threshold 14217), anl: -1898 (threshold
7058)
DEBUG: tablename: vac: 9917 (threshold 14217), anl: -798 (threshold
7058)
DEBUG: tablename: vac: 10987 (threshold 14217), anl: 272 (threshold
7058)
DEBUG: tablename: vac: 12016 (threshold 14217), anl: 1301 (threshold
7058)
DEBUG: tablename: vac: 12929 (threshold 14217), anl: 2214 (threshold
7058)
DEBUG: tablename: vac: 13717 (threshold 14217), anl: 3002 (threshold
7058)
DEBUG: tablename: vac: 14441 (threshold 14217), anl: 3727 (threshold
7058)
...vacuum threshold is hit...
DEBUG: tablename: vac: 752 (threshold 14217), anl: -9962 (threshold
7058)
DEBUG: tablename: vac: 1491 (threshold 14217), anl: -9223 (threshold
7058)
DEBUG: tablename: vac: 2213 (threshold 14217), anl: -8501 (threshold
7058)
DEBUG: tablename: vac: 2984 (threshold 14217), anl: -7730 (threshold
7058)

The outcome of this is that an ANALYZE is never run, as after the
counters are reset for each VACUUM the counter for ANALYZE gets
increasingly larger. But as you can see from our entries in
pg_autovacuum above, an ANALYZE should occur much more frequently
than a VACUUM.

We're running PostgreSQL 8.1.4 on a RHEL-3 with a 2.4.21-27.0.2.ELsmp
kernel.

Input is appreciated to explain exactly what is happening here.
Thanks so much!
--
Dylan Hansen
Enterprise Systems Developer