benchmark q6 producing incorrect result

[andygrove]

Describe the bug

---- tests::verify_q6 stdout ----
Running benchmarks with the following options: DataFusionBenchmarkOpt { query: 6, debug: false, iterations: 1, partitions: 2, batch_size: 8192, path: "/home/andy/git/apache/arrow-datafusion/benchmarks/data", file_format: "tbl", mem_table: false, output_path: None, disable_statistics: false }
Query 6 iteration 0 took 2977.2 ms and returned 1 rows
Query 6 avg time: 2977.24 ms
thread 'tests::verify_q6' panicked at 'assertion failed: `(left == right)`
  left: `["123141078.23"]`,
 right: `["75207429.08"]`', benchmarks/src/bin/tpch.rs:1004:13

To Reproduce
Steps to reproduce the behavior:

Expected behavior
A clear and concise description of what you expected to happen.

Additional context
Add any other context about the problem here.

[andygrove]

This issue is not related to decimal support. I see the same behavior after reverting to use floats.

Expected: 123141078.23; Actual: 75207768.18549997; Tolerance: 1

[andygrove]

If I change part of the query from this:

and l_discount between 0.06 - 0.01 and 0.06 + 0.01

to this:

and l_discount between 0.05 and 0.07

then it works

[andygrove]

never mind, this is related to decimal so we'll have to wait until decimals are supported to fix this one

this is how the query is running in the physical plan and this is a different range of rows compares to between 0.05 and 0.07.

and l_discount between 0.049999999 and 0.069999999

[andygrove]

These floats are being cast to decimal, which is not what we want:

CAST(lineitem.l_discount AS Decimal128(30, 15))lineitem.l_discount >= Decimal128(Some(49999999999999),30,15) 
AND 
CAST(lineitem.l_discount AS Decimal128(30, 15))lineitem.l_discount <= Decimal128(Some(69999999999999),30,15)

[viirya]

If I change part of the query from this:

and l_discount between 0.06 - 0.01 and 0.06 + 0.01

to this:

and l_discount between 0.05 and 0.07

then it works

Does it work? I just tried but it still doesn't work. The result 123140554.79.

[andygrove]

Does it work? I just tried but it still doesn't work. The result 123140554.79.

You are correct, the result is still wrong, just closer:

  left: `["123141078.23"]`,
 right: `["123140554.79"]`', benchmarks/src/bin/tpch.rs:1003:13

[andygrove]

I am investigating more and there are potentially multiple issues here.

I have filed #4071 for the following issue that affects the filter:

I am comparing Spark and DataFusion behavior against a Parquet file:

Spark

scala> spark.read.parquet("/mnt/bigdata/tpch/sf1-parquet/lineitem").createTempView("lineitem")
                                                                                
scala> spark.sql("SELECT count(*) from lineitem where l_discount between 0.05 and 0.07").show
+--------+                                                                      
|count(1)|
+--------+
|16361562|
+--------+


scala> spark.sql("SELECT count(*) from lineitem where l_discount between 0.06-0.01 and 0.06+0.01").show
+--------+
|count(1)|
+--------+
|16361562|
+--------+

DataFusion

❯ create external table lineitem stored as parquet location '/mnt/bigdata/tpch/sf1-parquet/lineitem';
0 rows in set. Query took 0.015 seconds.

❯ select count(*) from lineitem where l_discount between 0.05 and 0.07;
+-----------------+
| COUNT(UInt8(1)) |
+-----------------+
| 16361562        |
+-----------------+
1 row in set. Query took 0.487 seconds.

❯ select count(*) from lineitem where l_discount between 0.06-0.01 and 0.06+0.01;
+-----------------+
| COUNT(UInt8(1)) |
+-----------------+
| 10908630        |
+-----------------+
1 row in set. Query took 0.394 seconds.

So between 0.05 and 0.07 is consistent between Spark and DataFusion at least

[viirya]

❯ create external table lineitem stored as parquet location '/mnt/bigdata/tpch/sf1-parquet/lineitem';
0 rows in set. Query took 0.015 seconds.

❯ select count(*) from lineitem where l_discount between 0.05 and 0.07;
+-----------------+
| COUNT(UInt8(1)) |
+-----------------+
| 16361562 |
+-----------------+
1 row in set. Query took 0.487 seconds.

❯ select count(*) from lineitem where l_discount between 0.06-0.01 and 0.06+0.01;
+-----------------+
| COUNT(UInt8(1)) |
+-----------------+
| 10908630 |
+-----------------+
1 row in set. Query took 0.394 seconds.
So between 0.05 and 0.07 is consistent between Spark and DataFusion, but between 0.06-0.01 and 0.06+0.01 is not.

This is understandable. This is because 0.06 is parsed as double type in DataFusion but as decimal type in Spark. 0.06 - 0.01 will result in 0.049999999 instead of 0.05 if they treated as double type literal (no matter DataFusion or Spark) due to precision of floating point value.

More interestingly, for the query select count(*) from lineitem where l_discount between 0.05 and 0.07;, I only got 1637557 while running the query on DataFusion with the data generated by ./tpch-gen.sh 1.0.

I assume the query execution is correct. Then I will suspect that if the data generated by tpch-gen.sh is the same as '/mnt/bigdata/tpch/sf1-parquet/lineitem'`?

I also suspect that 123140554.79 is correct answer for the query q6 on this data, instead of 123141078.23.