Re: Mysterious Cache Problem

[Rick James]

Are the "reads" plain SELECT? Or SELECT .. FOR UPDATE? Or SELECT .. LOCK IN SHARE MODE?

Is there a mixture of such in the same transaction?

In REPEATABLE_READ, this is possible (though arguably "wrong"):

BEGIN
SELECT -- get 1
SELECT FOR UPDATE -- get 1
((another connection updates the row))
SELECT FOR UPDATE -- get 2 -- not really "repeatable"
SELECT -- get 1 -- "repeatable"
COMMIT

Is that what you are seeing?

Suggest...
* Do not have long running transaction
* Do not mix SELECT types without understanding that the isolation modes get confused.