Unfortunately https://review.typo3.org/58951/ did not actually
solve issues #86028 and #86491 for good.
There are two issues concerning the former approach:
1) The expiration time of all created caches was 0, which resulted
in a permanent creation and deletion of cache entries. This
behaviour cannot be called caching.
2) Number 1) increases the chance for race conditions where several
parallel requests tried to create the same cache entry.
To fix this, the check for an existing cache entry will be reverted
to behave like before the regression, i.e. cache entries with an
expiration timestamp of 0 are considered valid again.
Also, new caches are created within a transaction, which prevents
duplicate key errors.