[
https://issues.apache.org/jira/browse/OOZIE-3387?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Andras Piros updated OOZIE-3387:
--------------------------------
Description:
During data input dependency check Oozie evaluates EL functions like
{{coord:latest()}} using a non-optimal way which may result more than necessary
HDFS URI checks.
1. If the {{dataset}} frequency does not match the {{uri-template}} it checks
the same HDFS URI multiple times. For instance in the following definition:
{noformat}
<dataset name="dataset1" frequency="${coord:minutes(1)}"
initial-instance="2017-01-01T08:15Z" timezone="UTC">
<uri-template>${nameNode}/${rootDir}/${YEAR}-${MONTH}-${DAY}</uri-template>
<done-flag>_SUCCESS</done-flag>
</dataset>
...
<data-in name="coordInput" dataset="dataset1">
<instance>${coord:latest(0)}</instance>
</data-in>
{noformat}
oozie check the same {{.../2018-11-20/_SUCCESS}} file 24*60=1440 times. It
would be enough to check the file only once and skip the other 1439 tests.
2. If the frequency is 1 day and {{uri-template}} is definied in the following
way:
{noformat}
<uri-template>${nameNode}/${rootDir}/${YEAR}/${MONTH}/${DAY}</uri-template>
{noformat}
oozie will check the following directories one by one even if the some of the
parent directories are missing:
{noformat}
2018/11/20
2018/11/19
2018/11/18
...
{noformat}
If there is no {{2018/11}} directory then it is not necessary to check all the
{{2018/11/xx}} directories. It would be possible to reduce the number of HDFS
URI checks.
was:
During data input dependency check Oozie evaluates EL functions like {{
coord:latest}} using a non-optimal way which may result more than necessary
HDFS URI checks.
1. If the {{dataset}} frequency does not match the {{uri-template}} it checks
the same HDFS URI multiple times. For instance in the following definition:
{noformat}
<dataset name="dataset1" frequency="${coord:minutes(1)}"
initial-instance="2017-01-01T08:15Z" timezone="UTC">
<uri-template>${nameNode}/${rootDir}/${YEAR}-${MONTH}-${DAY}</uri-template>
<done-flag>_SUCCESS</done-flag>
</dataset>
...
<data-in name="coordInput" dataset="dataset1">
<instance>${coord:latest(0)}</instance>
</data-in>
{noformat}
oozie check the same {{.../2018-11-20/_SUCCESS}} file 24*60=1440 times. It
would be enough to check the file only once and skip the other 1439 tests.
2. If the frequency is 1 day and {{uri-template}} is definied in the following
way:
{noformat}
<uri-template>${nameNode}/${rootDir}/${YEAR}/${MONTH}/${DAY}</uri-template>
{noformat}
oozie will check the following directories one by one even if the some of the
parent directories are missing:
{noformat}
2018/11/20
2018/11/19
2018/11/18
...
{noformat}
If there is no {{2018/11}} directory then it is not necessary to check all the
{{2018/11/xx}} directories. It would be possible to reduce the number of HDFS
URI checks.
> Optimize coordinator data input dependency search
> -------------------------------------------------
>
> Key: OOZIE-3387
> URL: https://issues.apache.org/jira/browse/OOZIE-3387
> Project: Oozie
> Issue Type: Improvement
> Affects Versions: 5.1.0
> Reporter: Andras Salamon
> Priority: Major
>
> During data input dependency check Oozie evaluates EL functions like
> {{coord:latest()}} using a non-optimal way which may result more than
> necessary HDFS URI checks.
> 1. If the {{dataset}} frequency does not match the {{uri-template}} it checks
> the same HDFS URI multiple times. For instance in the following definition:
> {noformat}
> <dataset name="dataset1" frequency="${coord:minutes(1)}"
> initial-instance="2017-01-01T08:15Z" timezone="UTC">
>
> <uri-template>${nameNode}/${rootDir}/${YEAR}-${MONTH}-${DAY}</uri-template>
> <done-flag>_SUCCESS</done-flag>
> </dataset>
> ...
> <data-in name="coordInput" dataset="dataset1">
> <instance>${coord:latest(0)}</instance>
> </data-in>
> {noformat}
> oozie check the same {{.../2018-11-20/_SUCCESS}} file 24*60=1440 times. It
> would be enough to check the file only once and skip the other 1439 tests.
> 2. If the frequency is 1 day and {{uri-template}} is definied in the
> following way:
> {noformat}
> <uri-template>${nameNode}/${rootDir}/${YEAR}/${MONTH}/${DAY}</uri-template>
> {noformat}
> oozie will check the following directories one by one even if the some of the
> parent directories are missing:
> {noformat}
> 2018/11/20
> 2018/11/19
> 2018/11/18
> ...
> {noformat}
> If there is no {{2018/11}} directory then it is not necessary to check all
> the {{2018/11/xx}} directories. It would be possible to reduce the number of
> HDFS URI checks.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
