In Bitbucket Server versions up to and including 4.9, when scm-cache notices that there is not much

this.minFreeSpaceBytes = 1024 * 1024 * 1024; // 1GB default

space left on the device that stores the scm-cache files, it disables caching of new requests completely with a log message of the form:

2015-10-21 00:17:23,571 WARN  [ssh-scm-request-handler] bamboo_user *1DR16XBx17x1271154x131 it0suz 172.24.36.105 SSH - git-upload-pack '/hc/qt-client-build-tools.git' c.a.util.contentcache.ContentCache content-cache@3646: caching has been temporarily disabled because there is not enough space left on /opt/j2ee/domains/atlassian.com/stash-prod/home/caches/scm/3646

Instead of bypassing the cache when the disk gets too full, the scm-cache should choose "cold" (that is, not recently accessed) entries to evict instead. This should provide better performance when an instance is running with its disk(s) very nearly full.