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`.
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.
mem tracker can be logically divided into 4 layers: 1)process 2)type 3)query/load/compation task etc. 4)exec node etc.
type includes
enum Type {
GLOBAL = 0, // Life cycle is the same as the process, e.g. Cache and default Orphan
QUERY = 1, // Count the memory consumption of all Query tasks.
LOAD = 2, // Count the memory consumption of all Load tasks.
COMPACTION = 3, // Count the memory consumption of all Base and Cumulative tasks.
SCHEMA_CHANGE = 4, // Count the memory consumption of all SchemaChange tasks.
CLONE = 5, // Count the memory consumption of all EngineCloneTask. Note: Memory that does not contain make/release snapshots.
BATCHLOAD = 6, // Count the memory consumption of all EngineBatchLoadTask.
CONSISTENCY = 7 // Count the memory consumption of all EngineChecksumTask.
}
Object pointers are no longer saved between each layer, and the values of process and each type are periodically aggregated.
other fix:
In [fix](memtracker) Fix transmit_tracker null pointer because phamp is not thread safe #13528, I tried to separate the memory that was manually abandoned in the query from the orphan mem tracker. But in the actual test, the accuracy of this part of the memory cannot be guaranteed, so put it back to the orphan mem tracker again.
PR(https://github.com/apache/doris/pull/13404) introduced that ParquetReader
will break up batch insertion when encountering null values, which leads to the bad performance
compared to OrcReader.
So this PR has pushed null map into decode function, reduce the time of virtual function call
when encountering null values.
Further more, reuse hdfsFS among file readers to reduce the time of building connection to hdfs.
1. Fix issue #13115
2. Modify the method of `get_next_block` or `GenericReader`, to return "read_rows" explicitly.
Some columns in block may not be filled in reader, if the first column is not filled, use `block->rows()` can not return real row numbers.
3. Add more checks for broker load test cases.
Add more detail profile for ParquetReader:
ParquetColumnReadTime: the total time of reading parquet columns
ParquetDecodeDictTime: time to parse dictionary page
ParquetDecodeHeaderTime: time to parse page header
ParquetDecodeLevelTime: time to parse page's definition/repetition level
ParquetDecodeValueTime: time to decode page data into doris column
ParquetDecompressCount: counter of decompressing page data
ParquetDecompressTime: time to decompress page data
ParquetParseMetaTime: time to parse parquet meta data
This change serves the following purposes:
1. use ScanPredicate instead of TCondition for external table, it can reuse old code branch.
2. simplify and delete some useless old code
3. use ColumnValueRange to save predicate
refactor some arguments for parquet reader
1. Add new parquet context to wrap reader arguments
2. Reduced some arguments for function call
Co-authored-by: jinzhe <jinzhe@selectdb.com>
Add some utils and provide the candidate row range (generated with skipped row range of each column)
to read for page index filter
this version support binary operator filter
todo:
- use context instead of structures in close()
- process complex type filter
- use this instead of row group minmax filter
- refactor _eval_binary() for row group filter and page index filter
Currently, Doris has a variety of readers for different file formats,
such as parquet reader, orc reader, csv reader, json reader and so on.
The interfaces of these readers are not unified, which makes it impossible to call them through a unified method.
In this PR, I added a `GenericReader` interface class, and other Readers will implement this interface class
to use the `get_next_block()` method.
This PR currently only modifies `arrow_reader` and `parquet reader`.
Other readers will be modified one by one in subsequent PRs.
Read and generate parquet array column.
When D=1, R=0, representing an empty array. Empty array is not a null value, so the NullMap for this row is false,
the offset for this row is [offset_start, offset_end) whose `offset_start == offset_end`,
and offset_end is the start offset of the next row, so there is no value in the nested primitive column.
When D=0, R=0, representing a null array, and the NullMap for this row is true.
Parse parquet data with dictionary encoding.
Using the PLAIN_DICTIONARY enum value is deprecated in the Parquet 2.0 specification.
Prefer using RLE_DICTIONARY in a data page and PLAIN in a dictionary page for Parquet 2.0+ files.
refer: https://github.com/apache/parquet-format/blob/master/Encodings.md
1. Spark can set the timestamp precision by the following configuration:
spark.sql.parquet.outputTimestampType = INT96(NANOS), TIMESTAMP_MICROS, TIMESTAMP_MILLIS
DATETIME V1 only keeps the second precision, DATETIME V2 keeps the microsecond precision.
2. If using DECIMAL V2, the BE saves the value as decimal128, and keeps the precision of decimal as (precision=27, scale=9). DECIMAL V3 can maintain the right precision of decimal
Currently we use rapidjson to parse json document, It's fast but not fast enough compare to simdjson.And I found that the simdjson has a parsing front-end called simdjson::ondemand which will parse json when accessing fields and could strip the field token from the original document, using this feature we could reduce the cost of string copy(eg. we convert everthing to a string literal in _write_data_to_column by sprintf, I saw a hotspot from the flamegrame in this function, using simdjson::to_json_string will strip the token(a string piece) which is std::string_view and this is exactly we need).And second in _set_column_value we could iterate through the json document by for (auto field: object_val) {xxx}, this is much faster than looking up a field by it's field name like objectValue.FindMember("k1").The third optimization is the at_pointer interface simdjson provided, this could directly get the json field from original document.
Two improvements have been added:
1. Translate parquet physical type into doris logical type.
2. Decode parquet column chunk into doris ColumnPtr, and add unit tests to show how to use related API.
# Proposed changes
Read and decode parquet physical type.
1. The encoding type of boolean is bit-packing, this PR introduces the implementation of bit-packing from Impala
2. Create a parquet including all the primitive types supported by hive
## Remaining Problems
1. At present, only physical types are decoded, and there is no corresponding and conversion methods with doris logical.
2. No parsing and processing Decimal type / Timestamp / Date.
3. Int_8 / Int_16 is stored as Int_32. How to resolve these types.
* [feature](planner): push limit to olapscan when meet sort.
* if olap_scan_node's sort_info is set, push sort_limit, read_orderby_key
and read_orderby_key_reverse for olap scanner
* There is a common query pattern to find latest time serials data.
eg. SELECT * from t_log WHERE t>t1 AND t<t2 ORDER BY t DESC LIMIT 100
If the ORDER BY columns is the prefix of the sort key of table, it can
be greatly optimized to read much fewer data instead of read all data
between t1 and t2.
By leveraging the same order of ORDER BY columns and sort key of table,
just read the LIMIT N rows for each related segment and merge N rows.
1. set read_orderby_key to true for read_params and _reader_context
if olap_scan_node's sort info is set.
2. set read_orderby_key_reverse to true for read_params and _reader_context
if is_asc_order is false.
3. rowset reader force merge read segments if read_orderby_key is true.
4. block reader and tablet reader force merge read rowsets if read_orderby_key is true.
5. for ORDER BY DESC, read and compare in reverse order
5.1 segment iterator read backward using a new BackwardBitmapRangeIterator and
reverse the result block before return to caller.
5.2 VCollectIterator::LevelIteratorComparator, VMergeIteratorContext return
opposite result for _is_reverse order in its compare function.
Co-authored-by: jackwener <jakevingoo@gmail.com>
Analyze schema elements in parquet FileMetaData, and generate the hierarchy of nested fields.
For exmpale:
1. primitive type
```
// thrift:
optional int32 <column-name>;
// sql definition:
<column-name> int32;
```
2. nested type
```
// thrift:
optional group <column-name> (LIST) {
repeated group bag {
optional group array_element (LIST) {
repeated group bag {
optional int32 array_element
}
}
}
}
// sql definition:
<column-name> array<array<int32>>
```
When a rowset includes multiple segments, segments rows will be merged in generic_iterator but merged_rows is not maintained. Compaction will failed in check_correctness.
Co-authored-by: yixiutt <yixiu@selectdb.com>
* [Refactor][Bug-Fix][Load Vec] Refactor code of basescanner and vjson/vparquet/vbroker scanner
1. fix bug of vjson scanner not support `range_from_file_path`
2. fix bug of vjson/vbrocker scanner core dump by src/dest slot nullable is different
3. fix bug of vparquest filter_block reference of column in not 1
4. refactor code to simple all the code
It only changed vectorized load, not original row based load.
Co-authored-by: lihaopeng <lihaopeng@baidu.com>
1. solved the previous delayed unit test file size is too large (1.7G+) and the unit test link time is too long problem problems
2. Unify all unit tests into one file to significantly reduce unit test execution time to less than 3 mins
3. temporarily disable stream_load_test.cpp, metrics_action_test.cpp, load_channel_mgr_test.cpp because it will re-implement part of the code and affect other tests
