add a way to reduce the number of relatedness values stored for cache

[lfoppiano]

After hammering the system for several hours, the cache containing the list of all calculated relatedness reach the heap memory limit, causing the GC to be triggered too often (high CPU).

We should find an efficient solution to invalidate older element in this cache in order to keep the memory footprint lower.

[kermitt2]

Relatedness might disappear in version 0.0.4 for a LSTM-based architecture.

[lfoppiano]

the issue affect any batch process running for more than a day, a relatively simple cache implementation can be used via Guava https://github.com/google/guava/wiki/CachesExplained#caches

[tantikristanti]

Regarding this issue, the concept of cache is implemented in another branch called "cacheGuava". The implementation is done in a Relatedness class.

[lfoppiano]

Implemented in branch: https://github.com/kermitt2/entity-fishing/tree/cacheGuava

Tests using entity-fishing-client-python (branch: multitasking) - if problems, use goFishing, processBatch.py (with 1, 5, 10 threads), one after the other:

• Pubmed central (1943 pdfs)
• corpus of pdfs we used to test (pdf directory)

Tests to run (3 time per test):

1. master version without cache
2. master version with cache (branch 'guavaCache')

[tantikristanti]

Here are some tests that have been done so far with GoFishing: