Push TopN through Join.
JoinType just can be left/right outer join or cross join, because data of their one child can't be filtered.
new TopN is (original limit + original offset, 0) as limit and offset.
`ExternalFileTableValuedFunction` now has 3 derived classes:
- LocalTableValuedFunction
- HdfsTableValuedFunction
- S3TableValuedFunction
All these tvfs are for reading data from file. The difference is where to read the file, eg, from HDFS or from local filesystem.
So I refine the fields and methods of these classes.
Now there 3 kinds of properties of these tvfs:
1. File format properties
File format properties, such as `format`, `column_separator`. For all these tvfs, they are common properties.
So these properties should be analyzed in parenet class `ExternalFileTableValuedFunction`.
2. URI or file path
The URI or file path property indicate the file location. For different storage, the format of the uri are not same.
So they should be analyzed in each derived classes.
3. Other properties
All other properties which are special for certain tvf.
So they should be analyzed in each derived classes.
There are 2 new classes:
- `FileFormatConstants`: Define some common property names or variables related to file format.
- `FileFormatUtils`: Define some util methods related to file format.
After this PR, if we want to add some common properties for all these tvfs, only need to handled it in
`ExternalFileTableValuedFunction`, to avoid missing handle it in any one of them.
### Behavior change
1. Remove `fs.defaultFS` property in `hdfs()`, it can be got from `uri`
2. Use `\t` as the default column separator of csv format, same as stream load
The current implementation needs to iterate all metrics in a lock,
which might cause latency spikes. This PR changes the underlying
data structure to ConcurrentHashMap so that removing metrics doesn't
need to block the entire registry.
BE will crash because done run twice when try_offer() failed in
request_slave_tablet_pull_rowset.
Signed-off-by: freemandealer <freeman.zhang1992@gmail.com>
when we do NormalizeToSlot, we pushed complex expression and only remain
slot of it. When we do this, we collect alias and their child and
compute its child in bottom project, remain the result slot in current
node. for example
Window(max(...), c1 as a1)
after normalization, we get
Window(max(...), a1)
+-- Project(..., c1 as a1)
But, in some cases, we remove some SlotReference by mistake, for example
Window(max(...), c1, c1 as a1)
after normalization, we get
Window(max(...), a1)
+-- Project(..., c1 as a1)
we lost the SlotReference c1. This PR fix this problem. After this Pr,
we get
Window(max(...), c1, a1)
+-- Project(..., c1, c1 as a1)
Current multi-window plan generation has problem on the project sequence, for example:
+--LogicalWindow ( windowExpressions=[avg(sum_sales#115) WindowSpec(...) AS `avg_monthly_sales`#116, rank() WindowSpec(...) AS `rn`#117], ...)
and correspond physical plan is:
+--PhysicalWindow[6572]@16 ( windowFrameGroup=(Funcs=[avg(sum_sales#115) WindowSpec(...) AS `avg_monthly_sales`#116], ... )
+--PhysicalWindow[6568]@29 ( windowFrameGroup=(Funcs=[rank() WindowSpec(...) AS `rn`#117], ...] )
If the final plan is generated as following:
MultiCastDataSinks
STREAM DATA SINK
EXCHANGE ID: 20
HASH_PARTITIONED: rn[#208], i_brand[#202], cc_name[#203], i_category[#201]
Before we eventually resolve the multi-window issue, we add a projection as following and force a mapping but this will not cover all potential problems.
MultiCastDataSinks
STREAM DATA SINK
EXCHANGE ID: 20
HASH_PARTITIONED: rn[#219], i_brand[#213], cc_name[#214], i_category[#212]
PROJECTIONS: i_category[#184], i_brand[#185], cc_name[#186], d_year[#187], d_moy[#188], sum_sales[#189], avg_monthly_sales[#191], rn[#190]
PROJECTION TUPLE: 20
