Don,

1. Wasn't it performed mass delete or update on the indexed columns
when IB bevaiour is as you described? I mean mass is millions of raws.
If so, it is like on garbage collection on disconnect.
2. Do you make sweep manually from time to time? And don't some of
your users left started transaction forever and go home? And don't you
overuse CommitRetaining? Very large difference beetwen current
transaction number and oldest active or oldest rolled back can cause
dramatically effects.
3. Are you shure there is'nt light corruption in your database? Does
backup/restore normalize behaviour for a long period? Had you checked
database with gfix in read-only mode?
4. I remember some discussions about such behavior of Borland IB6
superserver where reason was not defined. Can't remember such
discussions about Firebird. Maybe there was some undefined bug in this
old engine that was fixed in Firebird.

Best regards, Alexander V.Nevsky.