Yibo Dong created SPARK-61529:
---------------------------------
Summary: approx_count_distinct undercounts arrays with trailing
NULL elements
Key: SPARK-61529
URL: https://issues.apache.org/jira/browse/SPARK-61529
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 4.3.0
Environment: {code}
Spark: 4.3.0
Commit: 98bc0212c94705434471627068f41e374ccab143
Java: OpenJDK 17.0.17
OS: macOS 15.7.4
Master: local[2]
{code}
Reporter: Yibo Dong
{{approx_count_distinct}} substantially underestimates the cardinality of
arrays that differ only by a trailing NULL element.
For example, {{[0]}} and {{[0, NULL]}} are distinct SQL array values but
receive the same {{xxhash64}} value.
When constructing 100 pairs of {{[i]}} and {{[i, NULL]}}, the input contains
200 distinct arrays. However, {{approx_count_distinct}} estimates only 99 at
the default relative standard deviation and 100 with {{relativeSD = 0.01}}.
h2. Steps to Reproduce
Save the following SQL as {{tmp.sql}}:
{code}
WITH ids AS (
SELECT id FROM range(100)
),
input AS (
SELECT array(CAST(id AS INT)) AS a FROM ids
UNION ALL
SELECT array(CAST(id AS INT), CAST(NULL AS INT)) AS a FROM ids
)
SELECT
count(DISTINCT a) AS exact_count,
approx_count_distinct(a) AS default_estimate,
approx_count_distinct(a, 0.01) AS tighter_estimate,
count(DISTINCT xxhash64(a)) AS hash_count
FROM input;
{code}
Execute:
{code}
spark-sql --master 'local[2]'
--conf spark.ui.enabled=false -f tmp.sql
{code}
h2. Expected Result
The input contains 200 distinct arrays. Both approximate aggregates should
estimate a cardinality near 200, subject to their configured approximation
errors.
The {{relativeSD = 0.01}} configuration specifies a relative standard deviation
of 1%, making an estimate near 100 especially inconsistent with the input
cardinality.
h2. Actual Result
{code}
exact_count | default_estimate | tighter_estimate | hash_count
------------+------------------+------------------+-----------
200 | 99 | 100 | 100
{code}
Both approximate aggregates underestimate the distinct count by approximately
50%, even with the tighter accuracy setting.
The hash-count result also shows that the 200 distinct arrays produce only 100
distinct {{xxhash64}} values. Each {{[i]}} and {{[i, NULL]}} pair has the same
hash, demonstrating a systematic collision pattern.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]