Hi Shrijeet, you have to distinguish between the storage format and the client side objects. KeyValue is an outlier (of sorts) as it is used on both server and client). Timestamps are per cell (KeyValue).
A Put object is something you create on the client to describe a put operation to be performed at the server. The server will take the information from the Put and write the necessary KeyValues into the Memstore (which will eventually be flushed to disk). So if you trigger a Put operations from the client and you change (say) 3 columns, the server will insert 3 KeyValues into the Memstore all of which carry the TS of the Put. Having the TS per cell (or KeyValue) is necessary to enforce ACID guarantees, which state that what you retrieve with Get is a set of KeyValues such as this combination of versions of KeyValues for this row existed together at a point. (need to remember here that multiple Put operations could insert different columns for the same rowKey). Makes sense? -- Lars ----- Original Message ----- From: Shrijeet Paliwal <[email protected]> To: [email protected]; lars hofhansl <[email protected]> Cc: Sent: Monday, November 28, 2011 4:31 PM Subject: Re: delete operation with timestamp Slightly offtopic, sorry. While we have attention on timestamps may I ask why HBase maintains a timestamp at row level (initialized with LATEST_TIMESTAMP)? In other words timestamp has meaning in context of a cell and HBase keeps it at that level, then why keep one TS at row level. Going further, what is the meaning of a timestamp 'ts' associated with Put object if all the KeyValue objects associated have timestamp different than 'ts'. Was the motivation behind this, to allow client not specify timestamp (in turn assume they meant latest ts)? I am looking at line 5 of this function http://pastebin.com/ik1Dxgqq which is serializing timestamp at row level and at lines 18-21 which are serializing timestamp at cell level. Thanks. On Mon, Nov 28, 2011 at 3:56 PM, lars hofhansl <[email protected]> wrote: > Hi Yi, > the reason is that nothing is ever changed in-place in HBase, only new files > are created (with the exception of the WAL, which is appended to, > and some special scenario like atomic increment and atomic appends, where > older version of the cells are removed from the memstore). > > That caters very well to the performance characteristics of the underlying > distributed file system (HDFS). > > > Consequently deleted rows are not actually deleted right away, we just record > the fact the rows should not be visible anymore and can eventually be removed. > The actual removal happens during the next compaction when new files are > created. > > Sometimes that does lead to unexpected behaviors such as the one you describe > below. > > In the trunk version of HBase I introduced the possibility to perform > time-range queries that can "peek" behind delete markers to retrieve cells > that are marked as deleted. (HBASE-4536) > > -- Lars > > > ----- Original Message ----- > From: Yi Liang <[email protected]> > To: [email protected] > Cc: > Sent: Thursday, November 24, 2011 10:11 PM > Subject: Re: delete operation with timestamp > > Thanks Daniel for your explanation. But still curious why we do such > design, it's unexpected for me. > > Also, this behavior of deleteColumns make delete operation not very user > friendly, why not use deleteColumn instead in hbase shell and thrift client? > > Thanks, > Yi > > 2011/11/24 Daniel Gómez Ferro <[email protected]> > >> >> On Nov 24, 2011, at 08:38 , Yi Liang wrote: >> >> > We're using hbase-0.90.3 with thrift client, and have encountered some >> > problems when we want to delete one specific version of a cell. >> > >> > First, there's no corresponding thrift api for Delete#deleteColumn(byte >> [] >> > family, byte [] qualifier, long timestamp). Instead, deleteColumns is >> > supported in mutateRowTs. But what we want is deleteColumn as we need to >> > keep the older versions. IMO, we should implement mutateRowTs >> > with deleteColumn, rather than deleteColumns. The hbase shell's delete >> > command has the same problem. >> > >> > Second, we find we can't reinsert any older cell if we have deleted that >> > cell with deleteColumns. For example: >> > hbase(main):007:0> scan 'test3' >> > ROW COLUMN+CELL >> > 0 row(s) in 0.0110 seconds >> > >> > hbase(main):008:0> put 'test3', 'r1', 'f1:c1', 'old', 1315550678308 >> > 0 row(s) in 0.0100 seconds >> > >> > hbase(main):009:0> scan 'test3' >> > ROW COLUMN+CELL >> > r1 column=f1:c1, >> > timestamp=1315550678308, value=old >> > 1 row(s) in 0.0290 seconds >> > >> > hbase(main):012:0> put 'test3', 'r1', 'f1:c1', 'new' >> > 0 row(s) in 0.0090 seconds >> > >> > hbase(main):013:0> scan 'test3' >> > ROW COLUMN+CELL >> > r1 column=f1:c1, >> > timestamp=1322119570316, value=new >> > 1 row(s) in 0.0140 seconds >> > >> > hbase(main):014:0> delete 'test3', 'r1', 'f1:c1', 1322119570316 >> > 0 row(s) in 0.0130 seconds >> > >> > hbase(main):015:0> scan 'test3' >> > ROW COLUMN+CELL >> > 0 row(s) in 0.0120 seconds >> > >> > hbase(main):016:0> put 'test3', 'r1', 'f1:c1', 'old', 1315550678308 >> > 0 row(s) in 0.0090 seconds >> > >> > hbase(main):017:0> scan 'test3' >> > ROW COLUMN+CELL >> > 0 row(s) in 0.0110 seconds >> > >> > There's no error message when we reinsert the old version, so we think it >> > has succeeded, but actually it's not. It looks like a bug. >> > >> > What's your opinion? >> > >> >> Hi, >> >> The second point is not a bug, it's how HBase is designed. Any delete >> (except deleteColumn) inserts a tombstone marker which masks any older >> value, so even if you insert later an older value it will be masked by the >> tombstone. You can see some nice examples here: >> http://outerthought.org/blog/417-ot.html >> >> There is also a new feature in trunk that allows you to retrieve masked >> values through a "raw scan" or a get with a timeRange that excludes the >> delete: https://issues.apache.org/jira/browse/HBASE-4536 >> >> Daniel >> >> > Thanks, >> > Yi >> >> > >
