fix: subscriber list cache size

[dunglas]

Fix a bug introduced in #1090. The subscriber list cache size must have a limit to prevent a leak.

May fix #1024.

cc @naxo8628, this may dramatically reduce memory usage.

[naxo8628]

@dunglas deployed & monitoring

[naxo8628]

First 6 hours seems similiar as the last deploy PR1092:

[dunglas]

Could you provide some profiles and maybe try to reduce the cache size using the new config parameter?

[naxo8628]

@dunglas
heap:
pprof.caddy.alloc_objects.alloc_space.inuse_objects.inuse_space.005.pb.gz

allocs:
pprof.caddy.alloc_objects.alloc_space.inuse_objects.inuse_space.006.pb.gz

tomorrow i'll test subscriber_list_cache_size with no default values

[naxo8628]

@dunglas
FYI:
12hs running fails on "Container mercure failed liveness probe, will be restarted"

settings:
topic_selector_cache 2000 128
subscriber_list_cache_size 10000

redeployed now with lower cache settings:
topic_selector_cache 1000 128
subscriber_list_cache_size 1000

[naxo8628]

Last 24h :

config:
topic_selector_cache 1000 128
subscriber_list_cache_size 1000

[dunglas]

@naxo8628 thanks. If I interpret correctly the graph, the problem looks fixed by this PR, right?

[naxo8628]

@dunglas yes. The memory/CPU is low and stable now! 😄

Last 48h: