Hi, Zhang Yang,
I've implemented the multi-version feature in my own Phoenix branch. But
this implementation is supposed to be working in a very very limited scenario
because there were so many things to think about when designing it. Here are
some primary problems that we must solve:
* add new syntax to support select with timestamps, we should support select
only one version and multi version within a range and the number of versions
too. For example:
select * from test timestamps min, max; // select all versions within
the specified time range
select * from test timestamps ts; // select a specified version
select * from test version number; // select specified number of
versions
select * from test version number timestamps min, max; // select specified
number of versions with a specified time range.
Note that this is not standard SQL syntax, which is not recommended.
* Timestamp is a Cell-level property in HBase, so we should support the same
thing in Phoenix. But how can we allow different timestamps for different
columns in the same row? I modified the ResultSet class and add some methods
like 'public Map<Long, T> getAllT(index)' to return all selected versions for a
single column. One can call this method on different columns for the same row
to retrieve all the things he wants. Users must use PhoenixResultSet instead of
ResultSet, this is not recommended either.
* How do we handle index updates/selects for multi-version? This is a messy
problem, so my implementation did not support multi-version for index tables.
* do not support GROUP BY, ORDER BY or any nested query/upsert.
* for batch commit, when you upsert the same row with different timestamps,
Phoenix can only commit the last timestamps you set. This is meaningless to do
this. So I simply forbid this scenario.
* Phoenix encoded the KVs into one Cell at the RS side, but if we want to
return multi-versions for different columns, especially different timestamps
for different columns, we must not do the encoding. So we must modify the
internals of Phoenix to support a brand new read path to do this.
Besides the huge efforts of implementing, IMHO, the primary problem is it's
not easy to implement this feature properly, as each one may have a different
requirement. You can implementing this feature personally in your personal
branch, but i don't know the best way to support this in an official Phoenix
release. What do you think of this? Any suggested design?
Thanks.
William.
At 2016-10-13 18:12:56, "Yang Zhang" <[email protected]> wrote:
Hello everyone
I saw that we can create a Phoenix table from an exist HBase table,(for detail)
My question is whether Phoenix can supprort the history version of my row?
I am trying to use Phoenix to store some info which have a lot of common
columns,
such as a table "T1 ( c1, c2, c3, c4 )", many rows share the same c1,c2,c3,and
the variable column is c4,
Using HBase we can put 'T1', 'key1', ' f:c4', 'new value', timestamp,
And i can get previous version of this row,They all share the same c1,c2,c3
whice HBase only store once.
Whether phoenix support to query history version of my row?
I got this jira link , This is same as my question.
Hadoop is using for big data, and mlutiple version can help us reduce our date
that unnecessary
I think phoenix should support this feature too.
If Phoenix shouldn't support multiple version, please tell me the reason.
Anyway thansks for your help, First
