Fix tow bugs:
1. Enabling file caching requires both `FE session` and `BE` configurations(enable_file_cache=true) to be enabled.
2. `ParquetReader` has not used `IOContext` previously, but `CachedRemoteFileReader::read_at` needs `IOContext` after PR(#17586).
Problem:
1. FE will split the parquet file into split. So a file can have several splits.
2. BE will scan each split, read the footer of the parquet file.
3. If 2 splits belongs to a same parquet file, the footer of this file will be read twice.
This PR mainly changes:
1. Use kv cache to cache the footer of parquet file.
2. The kv cache is belong to a scan node, so all parquet reader belong to this scan node will share same kv cache.
3. In cache, the key is "meta_file_path", the value is parsed thrift footer.
The KV Cache is sharded into mutlti sub cache.
So that different file can use different sub cache, avoid blocking each other
In my test, a query with 26 splits can reduce the footer parse time from 4s -> 1s
See #17764 for details
I have tested:
- Unit test for local/s3/hdfs/broker file system: be/test/io/fs/file_system_test.cpp
- Outfile to local/s3/hdfs/broker.
- Load from local/s3/hdfs/broker.
- Query file on local/s3/hdfs/broker file system, with table value function and catalog.
- Backup/Restore with local/s3/hdfs/broker file system
Not test:
- cold & host data separation case.
There are many type definitions in BE. Should unify the type system and simplify the development.
---------
Co-authored-by: yiguolei <yiguolei@gmail.com>
Support iceberg schema evolution for parquet file format.
Iceberg use unique id for each column to support schema evolution.
To support this feature in Doris, FE side need to get the current column id for each column and send the ids to be side.
Be read column id from parquet key_value_metadata, set the changed column name in Block to match the name in parquet file before reading data. And set the name back after reading data.
Since Filesystem inherited std::enable_shared_from_this , it is dangerous to create native point of FileSystem.
To avoid this behavior, making the constructor of XxxFileSystem a private method and using the static method create(...) to get a new FileSystem object.
The main purpose of this pr is to import `fileCache` for lakehouse reading remote files.
Use the local disk as the cache for reading remote file, so the next time this file is read,
the data can be obtained directly from the local disk.
In addition, this pr includes a few other minor changes
Import File Cache:
1. The imported `fileCache` is called `block_file_cache`, which uses lru replacement policy.
2. Implement a new FileRereader `CachedRemoteFilereader`, so that the logic of `file cache` is hidden under `CachedRemoteFilereader`.
Other changes:
1. Add a new interface `fs()` for `FileReader`.
2. `IOContext` adds some statistical information to count the situation of `FileCache`
Co-authored-by: Lightman <31928846+Lchangliang@users.noreply.github.com>
Fix three bugs when read iceberg v2 tables:
1. The `delete position` in `delete file` represents the position of delete row in the entire file, but the `read range` in
`RowGroupReader` represents the position in current row group. Therefore, we need to subtract the position of first
row of current row group from `delete position`.
2. When only reading the partition columns, `RowGroupReader` skips processing the `delete position`.
3. If the `delete position` has delete all rows in a row group, the `read range` is empty, but we read the whole row
group in such case.
Optimize four performance issues:
1. We change `delete position` to `delete range`, and then merge `delete range` and `read range` into the final read
ranges. This process is too tedious and time-consuming. . we can merge `delete position` and `read range` directly.
2. `delete position` is ordered in a `delete file`, so we can use merge-sort, instead of ordered-set.
3. Initialize `RowGroupReader` when reading, instead of initialize all row groups when opening a `ParquetReader`, to
save memory usage, and the same as `IcebergReader`.
4. Change the recursive call of `_do_lazy_read` to loop logic.
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
