The initial results of the EagerCacheSizeTest seem to indicate that we almost never benefit from the lazy permission evaluation (compared to reading all permission entries right away). From my understanding of the results the only exception are those cases where only very few items are being accessed (e.g. reading 100 items).
However, I am not totally sure if this is not a artifact of the random-read. I therefore started extending the benchmark with an option to re-read a randomly picked item more that once, which according to some analysis done quite some time ago is a common scenario specially when using Oak in combination with Apache Sling.
Benchmarks with 10-times re-reading the same random item:
As I would have expected it seems that the negative impact of lazy-loading is somewhat reduced, as the re-reading will hit the cache populated while reading.
Result are attached to
OAK-8662 (possibly more to come).