[
https://issues.apache.org/jira/browse/HBASE-15236?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15274797#comment-15274797
]
stack commented on HBASE-15236:
-------------------------------
We are careful about how we assign sequenceids. We do it in a single locality
intentionally to ensure our being able to assign edits a unique number. For
bulk loads currently, we do a little dance to get them a sequencid:
// We need to assign a sequential ID that's in between two memstores in
order to preserve
// the guarantee that all the edits lower than the highest sequential ID
from all the
// HFiles are flushed on disk. See HBASE-10958. The sequence id returned
when we flush is
// guaranteed to be one beyond the file made when we flushed (or if
nothing to flush, it is
// a sequence id that we can be sure is beyond the last hfile written).
if (assignSeqId) {
FlushResult fs = flushcache(true, false);
if (fs.isFlushSucceeded()) {
seqId = ((FlushResultImpl)fs).flushSequenceId;
} else if (fs.getResult() ==
FlushResult.Result.CANNOT_FLUSH_MEMSTORE_EMPTY) {
seqId = ((FlushResultImpl)fs).flushSequenceId;
} else {
throw new IOException("Could not bulk load with an assigned
sequential ID because the "+
"flush didn't run. Reason for not flushing: " +
((FlushResultImpl)fs).failureReason);
}
}
We'll flush current memstore just so we can insert the bulk loaded file with a
sequenceid that does not overlap with memstore content or with that of existing
hfiles. We do this crazy dance because all edits in the bulk loaded file get
this sequenceid and having no overlap makes it easy to know where it lies when
ordering for scans.
Is the issue that this sequenceid is lost on compactions after bulk load?
> Inconsistent cell reads over multiple bulk-loaded HFiles
> --------------------------------------------------------
>
> Key: HBASE-15236
> URL: https://issues.apache.org/jira/browse/HBASE-15236
> Project: HBase
> Issue Type: Bug
> Reporter: Appy
> Assignee: Appy
> Attachments: HBASE-15236-master-v2.patch,
> HBASE-15236-master-v3.patch, HBASE-15236-master-v4.patch,
> HBASE-15236-master-v5.patch, HBASE-15236-master-v6.patch,
> HBASE-15236-master-v7.patch, HBASE-15236-master-v8.patch,
> HBASE-15236-master-v9.patch, HBASE-15236.patch, TestWithSingleHRegion.java,
> tables_data.zip
>
>
> If there are two bulkloaded hfiles in a region with same seqID, same
> timestamps and duplicate keys*, get and scan may return different values for
> a key. Not sure how this would happen, but one of our customer uploaded a
> dataset with 2 files in a single region and both having same bulk load
> timestamp. These files are small ~50M (I couldn't find any setting for max
> file size that could lead to 2 files). The range of keys in two hfiles are
> overlapping to some extent, but not fully (so the two files are because of
> region merge).
> In such a case, depending on file sizes (used in
> StoreFile.Comparators.SEQ_ID), we may get different values for the same cell
> (say "r", "cf:50") depending on what we call: get "r" "cf:50" or get "r"
> "cf:".
> To replicate this, i ran ImportTsv twice with 2 different csv files but same
> timestamp (-Dimporttsv.timestamp=1). Collected two resulting hfiles in a
> single directory and loaded with LoadIncrementalHFiles. Here are the commands.
> {noformat}
> //CSV files should be in hdfs:///user/hbase/tmp
> sudo -u hbase hbase org.apache.hadoop.hbase.mapreduce.ImportTsv
> -Dimporttsv.columns=HBASE_ROW_KEY,c:1,c:2,c:3,c:4,c:5,c:6
> -Dimporttsv.separator='|' -Dimporttsv.timestamp=1
> -Dimporttsv.bulk.output=hdfs:///user/hbase/1 t tmp
> {noformat}
> tables_data.zip contains three tables: t, t2, and t3.
> To load these tables and test with them, use the attached
> TestWithSingleHRegion.java. (Change ROOT_DIR and TABLE_NAME variables
> appropriately)
> Hfiles for table 't':
> 1) 07922ac90208445883b2ddc89c424c89 : length = 1094: KVs = (c:5, 55) (c:6, 66)
> 2) 143137fa4fa84625b4e8d1d84a494f08 : length = 1192: KVs = (c:1, 1) (c:2, 2)
> (c:3, 3) (c:4, 4) (c:5, 5) (c:6, 6)
> Get returns 5 where as Scan returns 55.
> On the other hand if I make the first hfile, the one with values 55 and 66
> bigger in size than second hfile, then:
> Get returns 55 where as Scan returns 55.
> This can be seen in table 't2'.
> Weird things:
> 1. Get consistently returns values from larger hfile.
> 2. Scan consistently returns 55.
> Reason:
> Explanation 1: We add StoreFileScanners to heap by iterating over list of
> store files which is kept sorted in DefaultStoreFileManger using
> StoreFile.Comparators.SEQ_ID. It sorts the file first by increasing seq id,
> then by decreasing filesize. So our larger files sizes are always in the
> start.
> Explanation 2: In KeyValueHeap.next(), if both current and this.heap.peek()
> scanners (type is StoreFileScanner in this case) point to KVs which compare
> as equal, we put back the current scanner into heap, and call pollRealKV() to
> get new one. While inserting, KVScannerComparator is used which compares the
> two (one already in heap and the one being inserted) as equal and inserts it
> after the existing one. \[1\]
> Say s1 is scanner over first hfile and s2 over second.
> 1. We scan with s2 to get values 1, 2, 3, 4
> 2. see that both scanners have c:5 as next key, put current scanner s2 back
> in heap, take s1 out
> 4. return 55, advance s1 to key c:6
> 5. compare c:6 to c:5, put back s2 in heap since it has larger key, take out
> s1
> 6. ignore it's value (there is code for that too), advance s2 to c:6
> Go back to step 2 with both scanners having c:6 as next key
> This is wrong because if we add a key between c:4 and c:5 to first hfile, say
> (c:44, foo) which makes s1 as the 'current' scanner when we hit step 2, then
> we'll get the values - 1, 2, 3, 4, foo, 5, 6.
> (try table t3)
> Fix:
> Assign priority ids to StoreFileScanners when inserting them into heap,
> latest hfiles get higher priority, and use them in KVComparator instead of
> just seq id.
> \[1\]
> [PriorityQueue.siftUp()|http://grepcode.com/file/repository.grepcode.com/java/root/jdk/openjdk/6-b14/java/util/PriorityQueue.java#586]
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
