[SPARK-31430] Bug in the approximate quantile computation.

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 3.0.0
• Component/s: SQL
• Assignee: Unassigned
• Reporter: Siddartha Naidu
• Created: 13/Apr/20 05:02
• Updated: 09/Oct/20 18:44
• Resolved: 09/Oct/20 18:44

Description

I am seeing a bug where passing lower relative error to the approxQuantile function is leading to incorrect result in the presence of partitions. Setting a relative error 1e-6 causes it to compute equal values for 0.9 and 1.0 quantiles. Coalescing it back to 1 partition gives correct results. This issue was not present in spark version 2.4.5, we noticed it when testing 3.0.0-preview.

>>> df = spark.read.csv('file:///tmp/approx_quantile_data.csv', header=True, schema=T.StructType([T.StructField('Store',T.StringType(),True),T.StructField('seconds',T.LongType(),True)]))
>>> df = df.repartition(200, 'Store').localCheckpoint()
>>> df.approxQuantile('seconds', [0.8, 0.9, 1.0], 0.0001)
[1422576000.0, 1430352000.0, 1438300800.0]
>>> df.approxQuantile('seconds', [0.8, 0.9, 1.0], 0.00001)
[1422576000.0, 1430524800.0, 1438300800.0]
>>> df.approxQuantile('seconds', [0.8, 0.9, 1.0], 0.000001)
[1422576000.0, 1438300800.0, 1438300800.0]
>>> df.coalesce(1).approxQuantile('seconds', [0.8, 0.9, 1.0], 0.000001)
[1422576000.0, 1430524800.0, 1438300800.0]

Issue Links

• duplicates: SPARK-32908 percentile_approx() returns incorrect results