A native parquet reader for primitive types

[copperybean]

Changelog category (leave one):

• Performance Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

A native parquet reader, which can read parquet binary to ClickHouse Columns directly. Now this feature can be activated by setting input_format_parquet_use_native_reader to true.

Currently, parquet file is read by arrow library, and it's read to arrow table first, and then copy the arrow table to ClickHouse Columns. There are some shortcomings in performance.

• 1st, it's a redundant operation of copying data from arrow table to ClickHouse Columns.
• 2nd, in arrow's implementation, repetition and definition levels are read to int16_t array first; then nullability related valid_bits_ is generated. While, the null_map and offsets (not included this time) can be generated directly when reading definition and repetition levels.
• 3rd, for some situation, the arrow's implementation is inefficient, such as DataTypeDecimal32 type and dictionary parquet row group.

This feature is first implemented in the product BMR of Baidu AI Cloud, which has been fully tested.

Performance Test

As a result, the perforation of current implementation is speedup obviously. To generate the test data, a parquet file src.parquet of TPCDS table store_sales with scale 5000 is used, there are 35207247 rows in this file. Next, the test data is generated with following query:

select
  ss_item_sk,
  ss_wholesale_cost,
  ss_sales_price * ss_item_sk * ss_customer_sk as ss_sales_price_,
  concat('test string with ', toString(ss_item_sk % 1000)) as dict_str,
  concat('test string with ', toString(ss_item_sk * ss_cdemo_sk * ss_hdemo_sk)) as normal_str
from file('<path>/src.parquet', 'Parquet', 'ss_item_sk Nullable(Int32), ss_customer_sk Nullable(Int32), ss_cdemo_sk Nullable(Int32), ss_hdemo_sk Nullable(Int32), ss_wholesale_cost Nullable(Decimal(7, 2)), ss_sales_price Nullable(Decimal(21,6))')
into outfile '<path>/test.parquet' format Parquet

Then the performance is tested by following command

clickhouse-local --query "select max(<field-name>) from file('<path>/test.parquet', 'Parquet')" \
  --max_parsing_threads --max_threads 1 --input_format_parquet_use_native_reader <flag>

For each field, two types tests are triggered with different input_format_parquet_use_native_reader setting, and single thread is used. The parquet reading duration is counted as commit log duration while reading parquet. The CPU model used by this test is Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz.

The test result is detailed in following table

Column	Current Duration	Previous Duration	SpeedUp
ss_item_sk Nullable(Int32)	117ms	141ms	1.2
ss_wholesale_cost Nullable(Decimal(7, 2))	341ms	1132ms	3.3
ss_sales_price_ Nullable(Decimal(21, 6))	720ms	1185ms	1.6
dict_str Nullable(String)	259ms	624ms	2.4
normal_str Nullable(String)	717ms	1154ms	1.6

Documentation entry for user-facing changes

• Documentation is written (mandatory for new features)

Information about CI checks: https://clickhouse.com/docs/en/development/continuous-integration/

[clickhouse-ci]

This is an automatic comment. The PR descriptions does not match the template.

Please, edit it accordingly.

The error is: More than one changelog category specified: 'Improvement', 'Performance Improvement'

[robot-clickhouse]

This is an automated comment for commit 3d7befe with description of existing statuses. It's updated for the latest CI running

❌ Click here to open a full report in a separate page

Check name	Description	Status
A Sync	There's no description for the check yet, please add it to tests/ci/ci_config.py:CHECK_DESCRIPTIONS	⏳ pending
CI running	A meta-check that indicates the running CI. Normally, it's in success or pending state. The failed status indicates some problems with the PR	⏳ pending
Integration tests	The integration tests report. In parenthesis the package type is given, and in square brackets are the optional part/total tests	❌ failure
Mergeable Check	Checks if all other necessary checks are successful	❌ failure
Stateless tests	Runs stateless functional tests for ClickHouse binaries built in various configurations -- release, debug, with sanitizers, etc	❌ failure

Successful checks

Check name	Description	Status
AST fuzzer	Runs randomly generated queries to catch program errors. The build type is optionally given in parenthesis. If it fails, ask a maintainer for help	✅ success
ClickBench	Runs [ClickBench](https://github.com/ClickHouse/ClickBench/) with instant-attach table	✅ success
ClickHouse build check	Builds ClickHouse in various configurations for use in further steps. You have to fix the builds that fail. Build logs often has enough information to fix the error, but you might have to reproduce the failure locally. The cmake options can be found in the build log, grepping for cmake. Use these options and follow the general build process	✅ success
Compatibility check	Checks that clickhouse binary runs on distributions with old libc versions. If it fails, ask a maintainer for help	✅ success
Docker keeper image	The check to build and optionally push the mentioned image to docker hub	✅ success
Docker server image	The check to build and optionally push the mentioned image to docker hub	✅ success
Docs check	Builds and tests the documentation	✅ success
Fast test	Normally this is the first check that is ran for a PR. It builds ClickHouse and runs most of stateless functional tests, omitting some. If it fails, further checks are not started until it is fixed. Look at the report to see which tests fail, then reproduce the failure locally as described here	✅ success
Flaky tests	Checks if new added or modified tests are flaky by running them repeatedly, in parallel, with more randomization. Functional tests are run 100 times with address sanitizer, and additional randomization of thread scheduling. Integrational tests are run up to 10 times. If at least once a new test has failed, or was too long, this check will be red. We don't allow flaky tests, read the doc	✅ success
Install packages	Checks that the built packages are installable in a clear environment	✅ success
PR Check	There's no description for the check yet, please add it to tests/ci/ci_config.py:CHECK_DESCRIPTIONS	✅ success
Performance Comparison	Measure changes in query performance. The performance test report is described in detail here. In square brackets are the optional part/total tests	✅ success
Stateful tests	Runs stateful functional tests for ClickHouse binaries built in various configurations -- release, debug, with sanitizers, etc	✅ success
Stress test	Runs stateless functional tests concurrently from several clients to detect concurrency-related errors	✅ success
Style check	Runs a set of checks to keep the code style clean. If some of tests failed, see the related log from the report	✅ success
Unit tests	Runs the unit tests for different release types	✅ success
Upgrade check	Runs stress tests on server version from last release and then tries to upgrade it to the version from the PR. It checks if the new server can successfully startup without any errors, crashes or sanitizer asserts	✅ success

[copperybean]

Well, I think these break tests are not caused by my commit. UnitTestsAsan and ASTFuzzerTestAsan have succeeded in former test. While the integration test is failed because of logical error.

Hope for further suggestion.

[alexey-milovidov]

Yes, we need to start the investigation, and then check and fix every failure one by one.

[copperybean]

Yes, we need to start the investigation, and then check and fix every failure one by one.

@alexey-milovidov
About the investigation, do you refer to the investigation about this PR, or just about the failure tests?
And, should I rebase to master and trigger the checks again?

[alexey-milovidov]

@copperybean, please resolve conflicts.

[al13n321]

Cool, thanks for working on this!

Lots of review comments, mostly unimportant. The only blockers are:

• Seems that TIMESTAMP types will be incorrectly interpreted as seconds, please test.
• The block_missing_values_ptr thing, please test. If it doesn't work, at least add a TODO so we don't forget.
• Pass file metadata to ParquetFileReader::Open to avoid re-reading it for each row group.

[al13n321]

As this comment says, block_missing_values_ptr is used for applying default expressions to omitted fields (here "omitted" means NULLs, I think).

Native reader doesn't populate block_missing_values_ptr, so it'll probably break default expressions. Please test this:

create table a (x Int64, y String default 'asd') engine Memory
insert into function file('t.parquet') select 1 as x, null::Nullable(String) as y
insert into a select * from file('t.parquet')
select * from a

   ┌─x─┬─y───┐
1. │ 1 │ asd │
   └───┴─────┘

If it doesn't work, either populate block_missing_values_ptr based on null masks or at least add a TODO comment about that.

[copperybean]

Sorry, I missed this feature. I would like to add a TODO comment, because the block-missing-values logic will be modified again when supporting nested type, so the it can be implemented later for all types.

[al13n321]

A simpler way to handle nulls is to read all non-null values into a column, then use IColumn::expand to insert default values according to null mask. A little slower because of copying.

I don't mind the current approach, just mentioning the alternative in case it turns out better for the general case of arbitrarily nested arrays, nullables, and tuples.

[copperybean]

OK, I'd prefer to keep current implementation currently, may be I will change the logic as your suggestion when supporting nested columns.

[al13n321]

Would it be easier to create a ColumnLowCardinality and call convertToFullColumn() on it?

[copperybean]

Well, it will be easier. But the current implementation is more efficient, even not very much. Because the size of new Column generated by convertToFullColumn is smaller than reading_rows_num, so it should be resized to reading_rows_num again.

By the way, I will implement the logic as your suggestion next time. For current codes, I'd prefer not to change it right now. Instead, I can add some comments.

[copperybean]

Lots of review comments, mostly unimportant. The only blockers are:

Thanks for the comments, I will fix them in this weekend.

[al13n321]

:) insert into function file('u2.parquet') select number as a from numbers(100000) settings output_format_parquet_use_custom_encoder=0

INSERT INTO FUNCTION file('u2.parquet')
SETTINGS output_format_parquet_use_custom_encoder = 0
SELECT number AS a
FROM numbers(100000)
SETTINGS output_format_parquet_use_custom_encoder = 0

Query id: 67ad48c3-7952-4bb4-b1e5-9fe2959be6ad

Ok.

0 rows in set. Elapsed: 0.089 sec. Processed 100.00 thousand rows, 800.00 KB (1.12 million rows/s., 8.97 MB/s.)
Peak memory usage: 10.07 KiB.

:) select * from file('u2.parquet') limit 10 settings input_format_parquet_use_native_reader=1

SELECT *
FROM file('u2.parquet')
LIMIT 10
SETTINGS input_format_parquet_use_native_reader = 1

Query id: 4a17db55-e81c-418d-8b06-5c8f19d330dc


Elapsed: 0.020 sec. 

Received exception from server (version 24.5.1):
Code: 722. DB::Exception: Received from localhost:9000. DB::Exception: Unsupported logical type: Int(bitWidth=64, isSigned=false) and physical type: INT64 for field ==a==, bit width: 64: (in file/uri /home/ubuntu/cluster/n1/store/user_files/u2.parquet): While executing ParquetBlockInputFormat: While executing File. ()

[al13n321]

Confusing error message when trying to read an array:

:) insert into function file('v.parquet') select [number] as a from numbers(100000) settings output_format_parquet_use_custom_encoder=0

INSERT INTO FUNCTION file('v.parquet')
SETTINGS output_format_parquet_use_custom_encoder = 0
SELECT [number] AS a
FROM numbers(100000)
SETTINGS output_format_parquet_use_custom_encoder = 0

Query id: 1016edab-8db2-41b6-b273-eb77ec069125

Ok.

0 rows in set. Elapsed: 0.184 sec. Processed 100.00 thousand rows, 800.00 KB (543.01 thousand rows/s., 4.34 MB/s.)
Peak memory usage: 12.12 KiB.

 :) select * from file('v.parquet') limit 10 settings input_format_parquet_use_native_reader=1

SELECT *
FROM file('v.parquet')
LIMIT 10
SETTINGS input_format_parquet_use_native_reader = 1

Query id: cca172ee-98d6-42e8-ac9a-4bc837e055ed


Elapsed: 0.019 sec. 

Received exception from server (version 24.5.1):
Code: 722. DB::Exception: Received from localhost:9000. DB::Exception: can not find column with name: a: (in file/uri /home/ubuntu/cluster/n1/store/user_files/v.parquet): While executing ParquetBlockInputFormat: While executing File. ()

[al13n321]

Fix for the error message:

--- a/src/Processors/Formats/Impl/Parquet/ParquetRecordReader.cpp
+++ b/src/Processors/Formats/Impl/Parquet/ParquetRecordReader.cpp
@@ -312,16 +312,28 @@ ParquetRecordReader::ParquetRecordReader(
 {
     log = &Poco::Logger::get("ParquetRecordReader");
 
+    std::unordered_map<String, parquet::schema::NodePtr> parquet_columns;
+    auto root = file_reader->metadata()->schema()->group_node();
+    for (int i = 0; i < root->field_count(); ++i)
+    {
+        auto & node = root->field(i);
+        parquet_columns.emplace(node->name(), node);
+    }
+
     parquet_col_indice.reserve(header.columns());
     column_readers.reserve(header.columns());
     for (const auto & col_with_name : header)
     {
-        auto idx = file_reader->metadata()->schema()->ColumnIndex(col_with_name.name);
-        if (idx < 0)
-        {
-            auto msg = PreformattedMessage::create("can not find column with name: {}", col_with_name.name);
-            throw Exception(std::move(msg), ErrorCodes::PARQUET_EXCEPTION);
-        }
+        auto it = parquet_columns.find(col_with_name.name);
+        if (it == parquet_columns.end())
+            throw Exception(ErrorCodes::PARQUET_EXCEPTION, "no column with '{}' in parquet file", col_with_name.name);
+
+        auto node = it->second;
+        if (!node->is_primitive())
+            throw Exception(ErrorCodes::NOT_IMPLEMENTED, "arrays and maps are not implemented in native parquet reader");
+
+        auto idx = file_reader->metadata()->schema()->ColumnIndex(*node);
+        chassert(idx >= 0);
         parquet_col_indice.push_back(idx);
     }
     if (reader_properties.pre_buffer())

[Clark0]

It appears that the Bool and FLBA (fixed string) types are not currently supported, Are there any plans to add support for these types in the future?

[Clark0]

Bool types are distinct due to their 1-bit width, while clickhouse uses a UInt8 column vector

[copperybean]

With current framework, supporting these types is not difficult. My plan is to support nested type after this PR, then support all parquet types and enable 02735_parquet_encoder test.

[copperybean]

It seems that the ClickHouse special build check is failed because of OOM?

[al13n321]

It's broken in master, waiting for #64204 (and maybe something else, that one only fixes one of the two OOMing builds).

[Algunenano]

This should go to 24.6

[Algunenano]

#64359

[zhanglistar]

What a great feature! @copperybean Will you support complex types? @al13n321 Do you have plan to support complex types to make native parquet reader production ready?

[baibaichen]

@copperybean it's cool feature

Do you have plans to support page index? In this case，we need to skip reading some rows.