[ https://issues.apache.org/jira/browse/HBASE-9669?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13782794#comment-13782794 ]
Liang Xie commented on HBASE-9669: ---------------------------------- there're two paths: read and write. for read, yes, the current Hfile contains bloom filter related blocks and then loaded into block cache(and here still have some details need to be clear, IIRC, for data index or meta/bloom filter data, the current off-heap impl still place them in heap probably considering perf, just place the data block into off-heap). for write, seems to me the "ByteBuffer bloom" from ByteBloomFilter could be a potential bomb if large than PretenureSizeThreshold sometimes. Anyway, it's just a raw thought right now, will put more stuff after diving into the details:) > off-heap BloomFilter > -------------------- > > Key: HBASE-9669 > URL: https://issues.apache.org/jira/browse/HBASE-9669 > Project: HBase > Issue Type: Wish > Components: util > Affects Versions: 0.98.0 > Reporter: Liang Xie > > After lru block off-heap, the bloom filter is becoming one of the largest > contributers on java heap. it should be reasonable to have a config to > off-heap the bloom filter to alleviate gc hurt. -- This message was sent by Atlassian JIRA (v6.1#6144)