Excess use of memory mappings in multiprocess reader-writer scenarios

[finnschiermer]

Expected results

Core should reclaim memory mappings close(r) to the time when they are no longer used.

Actual Results

In a multiprocess setting (or any setting with multiple DB objects involved), where one process (or DB) only reads while others write AND these writes cause file growth AND the reader continuously refreshes their realm, stale memory mappings on the reader is not released. This is because release of memory mappings is done during commit, and in the described case, the reader does not commit. Doh!

On 32 bit machines this can cause a crash due to lack of virtual address space.

Core version

Core version: Any Core from Core-6 and onwards.

Prior to Core version 11.4.0 this bug could also hit a single-process app if it waited and refreshed while the sync client did a large multi-commit merge of data

[finnschiermer]

This will be fixed by #5645, as it for the above scenario only cause allocation of the minimum number of mappings.