JoshRosen opened a new pull request #26993: [WIP][SPARK-30338][SQL] Avoid 
unnecessary InternalRow copies in ParquetRowConverter
URL: https://github.com/apache/spark/pull/26993
 
 
   ⚠️ 📣 This PR is a work in progress; I'm opening it early for discussion / 
feedback. I may continue to add additional unit test cases. I think there's 
also some existing-but-outdated code comments that I might want to clean up.
   
   <!--
   Thanks for sending a pull request!  Here are some tips for you:
     1. If this is your first time, please read our contributor guidelines: 
https://spark.apache.org/contributing.html
     2. Ensure you have added or run the appropriate tests for your PR: 
https://spark.apache.org/developer-tools.html
     3. If the PR is unfinished, add '[WIP]' in your PR title, e.g., 
'[WIP][SPARK-XXXX] Your PR title ...'.
     4. Be sure to keep the PR description updated to reflect all changes.
     5. Please write your PR title to summarize what this PR proposes.
     6. If possible, provide a concise example to reproduce the issue for a 
faster review.
   -->
   
   ### What changes were proposed in this pull request?
   <!--
   Please clarify what changes you are proposing. The purpose of this section 
is to outline the changes and how this PR fixes the issue. 
   If possible, please consider writing useful notes for better and faster 
reviews in your PR. See the examples below.
     1. If you refactor some codes with changing classes, showing the class 
hierarchy will help reviewers.
     2. If you fix some SQL features, you can provide some references of other 
DBMSes.
     3. If there is design documentation, please add the link.
     4. If there is a discussion in the mailing list, please add the link.
   -->
   
   This PR modifies `ParquetRowConverter` to remove unnecessary 
`InternalRow.copy()` calls for structs that are directly nested in other 
structs. 
   
   ### Why are the changes needed?
   
   These changes  can significantly improve performance when reading Parquet 
files that contain deeply-nested structs with many fields.
   
   The `ParquetRowConverter` uses per-field `Converter`s for handling 
individual fields. Internally, these converters may have mutable state and may 
return mutable objects. In most cases, each `converter` is only invoked once 
per Parquet record (this is true for top-level fields, for example). However, 
arrays and maps may call their child element converters multiple times per 
Parquet record: in these cases we must be careful to copy any mutable outputs 
returned by child converters.
   
   In the existing code, `InternalRow`s are copied whenever they are stored 
into _any_ parent container (not just maps and arrays). This copying can be 
especially expensive for deeply-nested fields, since a deep copy is performed 
at every level of nesting.
   
   This PR modifies the code to avoid copies for structs that are directly 
nested in structs; see inline code comments for an argument for why this is 
safe. 
   
   ### Does this PR introduce any user-facing change?
   <!--
   If yes, please clarify the previous behavior and the change this PR proposes 
- provide the console output, description and/or an example to show the 
behavior difference if possible.
   If no, write 'No'.
   -->
   
   No.
   
   ### How was this patch tested?
   <!--
   If tests were added, say they were added here. Please make sure to add some 
test cases that check the changes thoroughly including negative and positive 
cases if possible.
   If it was tested in a way different from regular unit tests, please clarify 
how you tested step by step, ideally copy and paste-able, so that other 
reviewers can test and check, and descendants can verify in the future.
   If tests were not added, please describe why they were not added and/or why 
it was difficult to add.
   -->
   
   **Correctness**:  I added new test cases to `ParquetIOSuite` to increase 
coverage of nested structs, including structs nested in arrays: previously this 
suite didn't test that case, so we used to lack mutation coverage of this 
`copy()` code (the suite's tests still passed if I incorrectly removed the 
`.copy()` in all cases).
   
   **Performance**: I put together a simple local benchmark demonstrating the 
performance problems:
   
   First, construct a nested schema:
   
   ```scala
     case class Inner(
       f1: Int,
       f2: Long,
       f3: String,
       f4: Int,
       f5: Long,
       f6: String,
       f7: Int,
       f8: Long,
       f9: String,
       f10: Int
     )
     
     case class Wrapper1(inner: Inner)
     case class Wrapper2(wrapper1: Wrapper1)
     case class Wrapper3(wrapper2: Wrapper2)
   ```
   
   `Wrapper3`'s schema looks like:
   
   ```
   root
    |-- wrapper2: struct (nullable = true)
    |    |-- wrapper1: struct (nullable = true)
    |    |    |-- inner: struct (nullable = true)
    |    |    |    |-- f1: integer (nullable = true)
    |    |    |    |-- f2: long (nullable = true)
    |    |    |    |-- f3: string (nullable = true)
    |    |    |    |-- f4: integer (nullable = true)
    |    |    |    |-- f5: long (nullable = true)
    |    |    |    |-- f6: string (nullable = true)
    |    |    |    |-- f7: integer (nullable = true)
    |    |    |    |-- f8: long (nullable = true)
    |    |    |    |-- f9: string (nullable = true)
    |    |    |    |-- f10: integer (nullable = true)
   ```
   
   Next, generate some fake data:
   
   ```scala
     val data = spark.range(1, 1000 * 1000 * 25, 1, 1).map { i =>
       Wrapper3(Wrapper2(Wrapper1(Inner(
         i.toInt,
         i * 2,
         (i * 3).toString,
         (i * 4).toInt,
         i * 5,
         (i * 6).toString,
         (i * 7).toInt,
         i * 8,
         (i * 9).toString,
         (i * 10).toInt
       ))))
     }
   
     data.write.mode("overwrite").parquet("/tmp/parquet-test")
   ```
   
   I then ran a simple benchmark consisting of
   
   ```
   spark.read.parquet("/tmp/parquet-test").selectExpr("hash(*)").rdd.count()
   ```
   
   where the `hash(*)` is designed to force decoding of all Parquet fields but 
avoids `RowEncoder` costs in the `.rdd.count()` stage.
   
   In the old code, expensive copying takes place at every level of nesting; 
this is apparent in the following flame graph:
   
   
![image](https://user-images.githubusercontent.com/50748/71389014-88a15380-25af-11ea-9537-3e87a2aef179.png)
   
   After this PR's changes, the above toy benchmark runs ~30% faster.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
[email protected]


With regards,
Apache Git Services

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to