[SPARK-5683] [SQL] Avoid multiple json generator created

[chenghao-intel]

No description provided.

[SparkQA]

Test build #27092 has started for PR 4468 at commit e556aba.

• This patch merges cleanly.

[SparkQA]

Test build #27093 has started for PR 4468 at commit aeb7801.

• This patch merges cleanly.

[SparkQA]

Test build #27092 has finished for PR 4468 at commit e556aba.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/27092/
Test PASSed.

[SparkQA]

Test build #27093 has finished for PR 4468 at commit aeb7801.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/27093/
Test PASSed.

[chenghao-intel]

cc @marmbrus @yhuai

[yhuai]

Should we expect any test result change after using setRootValueSeparator(null)?

[chenghao-intel]

A (single space) will be inserted between 2 json records (in string) by default;
Previously, we create thegeneratorobject for each of record in serialization, hence it's OK with default object separator.
However, in this PR, we are trying to using a singlegenerator for all of the records, that's why we need to set it as null, which means nothing will be inserted between records.

[yhuai]

@chenghao-intel Thank you for the PR. It will be great if you can also attach some performance results.

[chenghao-intel]

@yhuai, thanks for reviewing, in a local testing, it shows about 10-15% performance gain.

[yhuai]

LGTM

[chenghao-intel]

@marmbrus any more comment on this?

[marmbrus]

Thanks, merged to master and 1.3