Enis Soztutar created HBASE-13853:
-------------------------------------
Summary: ITBLL data loss in HBase-1.1
Key: HBASE-13853
URL: https://issues.apache.org/jira/browse/HBASE-13853
Project: HBase
Issue Type: Bug
Reporter: Enis Soztutar
Assignee: Enis Soztutar
Priority: Blocker
Fix For: 2.0.0, 1.1.1
After 2 days of log mining with ITBLL Search tool (thanks Stack, BTW for the
tool), the sequence of events are like this:
One of mappers from Search reportedly finds missing keys in some WAL file:
{code}
2015-06-07 06:49:24,573 INFO [main] org.apache.hadoop.mapred.MapTask:
Processing split:
hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/oldWALs/os-enis-dal-test-jun-4-4.openstacklocal%2C16020%2C1433636320201.default.1433637872643
(-9223372036854775808:9223372036854775807) length:132679657 2015-06-07
06:49:24,639 INFO [main] org.apache.hadoop.hbase.mapreduce.WALInputFormat:
Opening reader for
hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/oldWALs/os-enis-dal-test-jun-4-4.openstacklocal%2C16020%2C1433636320201.default.1433637872643
(-9223372036854775808:9223372036854775807) length:132679657
2015-06-07 06:50:16,383 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Loaded keys
to find: count=2384870 2015-06-07 06:50:16,607 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Found
cell=!+\xF1CB\x08\x13\xA0W\x94\xC4\x1C\xDA\x1D\x0D\xBC/meta:prev/1433637873293/Put/vlen=16/seqid=0
2015-06-07 06:50:16,663 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Found
cell=!+\xF1CB\x08\x13\xA0W\x94\xC4\x1C\xDA\x1D\x0D\xBC/meta:count/1433637873293/Put/vlen=8/seqid=0
2015-06-07 06:50:16,664 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Found
cell=!+\xF1CB\x08\x13\xA0W\x94\xC4\x1C\xDA\x1D\x0D\xBC/meta:client/1433637873293/Put/vlen=71/seqid=0
2015-06-07 06:50:16,671 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Found
cell=$\x1A\x99\x06\x86\xE7\x07\xA2\xA7\xB2\xFB\xCEP\x12"\x04/meta:prev/1433637873293/Put/vlen=16/seqid=0
2015-06-07 06:50:16,672 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Found
cell=$\x1A\x99\x06\x86\xE7\x07\xA2\xA7\xB2\xFB\xCEP\x12"\x04/meta:count/1433637873293/Put/vlen=8/seqid=0
2015-06-07 06:50:16,678 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Found
cell=$\x1A\x99\x06\x86\xE7\x07\xA2\xA7\xB2\xFB\xCEP\x12"\x04/meta:client/1433637873293/Put/vlen=71/seqid=0
2015-06-07 06:50:16,679 INFO [main]
org.apache.hadoop.hbase.test.IntegrationTestBigLinkedList$Search: Found
cell=\x1F\x80#\x95\xAE:i=)S\x01\x08
\xD6\xFF\xD5/meta:prev/1433637873293/Put/vlen=16/seqid=0
{code}
{{hlog -p}} confirms the missing keys are there in the file:
{code}
Sequence=7276 , region=b086e29f909c9790446cd457c1ea7674 at write timestamp=Sun
Jun 07 00:44:33 UTC 2015
row=!+\xF1CB\x08\x13\xA0W\x94\xC4\x1C\xDA\x1D\x0D\xBC, column=meta:prev
value: \x1B\xF5\xF3^\x8D\xB4\x85\xE3\xF4wS\x9A]\x0D\xABe
row=!+\xF1CB\x08\x13\xA0W\x94\xC4\x1C\xDA\x1D\x0D\xBC, column=meta:count
value: \x00\x00\x00\x00\x002\x87l
row=!+\xF1CB\x08\x13\xA0W\x94\xC4\x1C\xDA\x1D\x0D\xBC, column=meta:client
value: Job: job_1433466891829_0002 Task:
attempt_1433466891829_0002_m_000003_0
{code}
When the RS gets killing from CM, master does WAL splitting:
{code}
./hbase-hbase-regionserver-os-enis-dal-test-jun-4-2.log:2015-06-07 00:46:12,581
INFO [RS_LOG_REPLAY_OPS-os-enis-dal-test-jun-4-2:16020-0] wal.WALSplitter:
Processed 2971 edits across 4 regions; edits skipped=740; log
file=hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/WALs/os-enis-dal-test-jun-4-4.openstacklocal,16020,1433636320201-splitting/os-enis-dal-test-jun-4-4.openstacklocal%2C16020%2C1433636320201.default.1433637872643,
length=132679657, corrupted=false, progress failed=false
{code}
The edits with Sequence=7276 should be in this recovered.edits file:
{code}
2015-06-07 00:46:12,574 INFO [split-log-closeStream-2] wal.WALSplitter: Closed
wap
hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/data/default/IntegrationTestBigLinkedList/b086e29f909c9790446cd457c1ea7674/recovered.edits/0000000000000007276.temp
(wrote 739 edits in 3950ms)
2015-06-07 00:46:12,580 INFO [split-log-closeStream-2] wal.WALSplitter: Rename
hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/data/default/IntegrationTestBigLinkedList/b086e29f909c9790446cd457c1ea7674/recovered.edits/0000000000000007276.temp
to
hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/data/default/IntegrationTestBigLinkedList/b086e29f909c9790446cd457c1ea7674/recovered.edits/0000000000000008752
{code}
Notice that the file {{0000000000000008752}} should contain {{(wrote 739 edits
in 3950ms)}}.
When the region gets opened again, it replays the recovered.edits files:
{code}
2015-06-07 00:46:17,764 DEBUG [RS_OPEN_REGION-os-enis-dal-test-jun-4-5:16020-0]
regionserver.HRegion: Found 4 recovered edits file(s) under
hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/data/default/IntegrationTestBigLinkedList/b086e29f909c9790446cd457c1ea7674
{code}
However, it does not see ANY entry in the recovered.edits file
{{0000000000000008752}}:
{code}
2015-06-07 00:46:17,796 DEBUG [RS_OPEN_REGION-os-enis-dal-test-jun-4-5:16020-0]
regionserver.HRegion: Applied 0, skipped 0, firstSequenceIdInLog=-1,
maxSequenceIdInLog=-1,
path=hdfs://os-enis-dal-test-jun-4-1.openstacklocal:8020/apps/hbase/data/data/default/IntegrationTestBigLinkedList/b086e29f909c9790446cd457c1ea7674/recovered.edits/0000000000000008752
{code}
Looking at the length of the file, it is very small which means that the edits
are not written.
Now, If you are still with me, I think the reason why we do not write any edit
to {{0000000000000008752}} although we say we have written 739 edits is due to
HBASE-12405.
Specifically, we pass {{FSHLog.getEarliestMemstoreSeqNum() - 1}} as the stores
flush seqId, which gets saved in the master and used with {{LastSequenceId}}.
The WALSplitter skips the cells per store using {{filterCellByStore()}} which
is logical. However, the data loss happens, because
{{FSHLog.getEarliestMemstoreSeqNum()}} ignores flushing edits. For example, if
we have seqId = 10 to 100 as flushing, and seqId 101 in the new memstore, and
heartbeat to master during the flush will carry 101 as the earliest flush
seqId. if the flush does not complete (due to RS getting killed) this results
in incorrectly skipped edits.
The WAL replay code also uses this kind of filtering so this may also explain
some of the failures with DLR.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
