[HDFS-7900] ShortCircuitCache.replicaInfoMap keeps too many deleted file descriptor

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Duplicate
• Affects Version/s: 2.3.0
• Component/s: hdfs-client

Environment

hadoop cdh5 2.3.0 hbase 0.98

• Assignee: Unassigned
• Reporter: zhangshilong
• Due: 05/Mar/15
• Created: 06/Mar/15 10:00
• Updated: 12/May/15 19:24
• Resolved: 12/May/15 19:24

Description

I delete some hbase's files manually or use "rm -rf blk_xxxx" to delete the blockfile directly, but hbase keeps the file descriptor very long time.

I found these file descriptor may be kept in shortcircuitcache replicaMap，
but could not find when the file descriptor will be removed. replicaMap has no limits size for putting.

run： lsof -p pid |grep deleted
part of result：
lk_1102309377_28571078.meta (deleted)
java 8430 hbase 8537r REG 8,145 536870912 806553760 /search/hadoop08/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir61/blk_1102541663 (deleted)
java 8430 hbase 8540r REG 8,113 4194311 812434001 /search/hadoop06/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir62/subdir21/blk_1102524193_28785917.meta (deleted)
java 8430 hbase 8541r REG 8,65 536870912 813718517 /search/hadoop03/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir31/subdir14/blk_1102523618 (deleted)
java 8430 hbase 8542r REG 8,65 4194311 813718518 /search/hadoop03/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir31/subdir14/blk_1102523618_28785342.meta (deleted)
java 8430 hbase 8543r REG 8,193 536870912 1886733815 /search/hadoop12/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir20/subdir22/blk_1102533549 (deleted)
java 8430 hbase 8544r REG 8,65 4194311 814828988 /search/hadoop03/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir49/blk_1102676585_28938309.meta (deleted)
java 8430 hbase 8545r REG 8,17 4194311 812962137 /search/hadoop10/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir53/blk_1102597493_28859217.meta (deleted)
java 8430 hbase 8546r REG 8,97 4194311 810468992 /search/hadoop05/data/current/BP-715213703-10.141.46.46-1418959337587/current/finalized/subdir4/subdir46/blk_1102524567_28786291.meta (deleted)

Issue Links

• duplicates: HDFS-6750 The DataNode should use its shared memory segment to mark short-circuit replicas that have been unlinked as stale