[KUDU-3054] Init kudu.write_duration accumulator lazily

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 1.9.0
• Fix Version/s: NA
• Component/s: spark
• Assignee: Unassigned
• Reporter: liupengcheng
• Created: 14/Feb/20 02:02
• Updated: 03/Jun/20 16:06
• Resolved: 03/Jun/20 16:06

Description

Currently, we encountered a issue in kudu-spark that will causing spark sql query failure:

```

Job aborted due to stage failure: Total size of serialized results of 942 tasks (2.0 GB) is bigger than spark.driver.maxResultSize (2.0 GB)

```

After carefully debug, we find out that it's the kudu.write_duration accumulators causing single spark task larger than 2M, thus all tasks size of the stage will bigger than the limit.

However, this stage is just reading kudu table and do shuffle exchange, no writing any kudu tables.

So I think should init this accumulator lazily in KuduContext to avoid such issues.

 

Issue Links

• duplicates: KUDU-3056 kudu-spark HdrHistogramAccumulator is too big, and make spark job failed
• links to: GitHub Pull Request #27