In previous implementation, when doing list partition prune, we need to generation `rangeToId`
every time we doing prune.
But `rangeToId` is actually a static data that should be create-once-use-every-where.
So for hive partition, I created the `rangeToId` and all other necessary data structures for partition prunning
in partition cache, so that we can use it directly.
In my test, the cost of partition prune for 10000 partitions reduce from 8s -> 0.2s.
Aslo add "partition" info in explain string for hive table.
```
| 0:VEXTERNAL_FILE_SCAN_NODE |
| predicates: `nation` = '0024c95b' |
| inputSplitNum=1, totalFileSize=4750, scanRanges=1 |
| partition=1/10000 |
| numNodes=1 |
| limit: 10 |
```
Bug fix:
1. Fix bug that es scan node can not filter data
2. Fix bug that query es with predicate like `where substring(test2,2) = "ext2";` will fail at planner phase.
`Unexpected exception: org.apache.doris.analysis.FunctionCallExpr cannot be cast to org.apache.doris.analysis.SlotRef`
TODO:
1. Some problem when quering es version 8: ` Unexpected exception: Index: 0, Size: 0`, will be fixed later.
PR https://github.com/apache/doris/pull/13917 has supported lazy read for non-predicate columns in ParquetReader,
but can't trigger lazy read when predicate columns are partition or missing columns.
This PR support such case, and fill partition and missing columns in `FileReader`.
Before this pr, if we try to load ORC file with native list(or array) type data, the be will crash.
Because complex types in ORC file include multi real columns, so we need to filter columns by column names.
Otherwise we could not read all columns we need.
Now arrow release-7.0.0 only support create stripe reader by column index, so we patch it to support create stripe reader by column names.
Co-authored-by: cambyzju <zhuxiaoli01@baidu.com>
In concurrent load, some publish timeout happens occasionally. This is
cause by meta lock hold by other thread so publish add increase rowset
hang for several seconds.
StorageEngine::start_delete_unused_rowset will hold gc_mutex and it cost
a lot of time, so that add_used_rowset wait lock, and compaction modify_rowset
or other tablet method will hold meta_lock and call add_unused_rowset which
will make meta_lock occupied for too long, finally makes publish timeout.
In this pr, I copy unused_rowsets in lock and delete these rowset without lock,
makes gc_mutex more lightweight so meta lock can be acquired immediately in publish thread.
My test shows that no publish timeout in concurrent stream load.
boost::stacktrace::stacktrace() has memory leak, so use glog internal func to print stacktrace.
The reason for the memory leak of boost::stacktrace is that a state is saved in the thread local of each thread but not actively released. The test found that each thread leaked about 100M after calling boost::stacktrace.
refer to:
boostorg/stacktrace#118boostorg/stacktrace#111
Fix three bugs:
1. The EOF of lazy read columns may be not equal to the EOF of predicate columns.
(for example: If the predicate column has 3 pages, with 400 rows for each, but the last page
is filtered by page index. When batch_size=992, the EOF of predicate column is true.
However, we should set batch_size=800 for lazy read column, so the EOF of lazy read column may be false.)
2. The array column does not count the number of nulls
3. Generate wrong NullMap for array column
When calling select on remote files, download cache files to local disk.
When calling alter table on remote files, read files directly from remote storage. So if tablet is too large, it will not take up too many local disk when creating local cache file.
For runtime filter, signal will be called by a thread which is different from the await thread. So there will be a potential race for variable is_ready
Currently, it takes too much time to build BE from source in workflow environments (P0/P1) which affects the efficiency of daily development.
We can measure the time by executing the following command.
time EXTRA_CXX_FLAGS='-O3' BUILD_TYPE=ASAN ./build.sh --be --fe --clean -j "$(nproc)"
This PR optimizes the compilation time by exploiting the following methods.
Reduce the codegen by removing some useless std::visit.
Disable the optimization for some template functions which are instantiated by std::visit conditionally (except for the RELEASE build).
When upgrade from 1.1 to master, and then rollback to 1.1, and upgrade to master again, BE will coredump because some rowsets has schema and some rowsets has no schema. In the first time upgrade from 1.1, BE will flush schema in all rowsets and after rollback to 1.1, BE do compaction, and create some new rowset without schema. And the second time upgrade from 1.1, BE coredump because some conditions depend on having all or none of the rowsets.
Read predicate columns firstly, and use VExprContext(push-down predicates)
to generate the select vector, which is then applied to read the non-predicate columns.
The data in non-predicate columns may be skipped by select vector, so the value-decode-time can be reduced.
If a whole page can be skipped, the decompress-time can also be reduced.
