7/11/2018 5:48:26 PM
active (#1824)
FlowRunnerManager has a variable named "installedProjects" to keep the projects currently residing in the shared project dir. It gets initialized when FlowRunnerManager is constructed and used to look up the project directory by project ID and version. It's maintained by a cleaner thread inside FlowRunnerManager and FlowPreparer - when new project is downloaded or old project is deleted, its corresponding in-memory mapping is populated or kicked out from installedProjects.
However current logic suffers from risk of cache out of sync. E.g, when FlowRunnerManager populates installedProjects upon construction, sometimes executor is still in inactive mode, bearing some verification work like running integration test. However, another executor at same host which is active is performing project cleaning up. Some projects might be deleted by the active executor but not reflected by installedProjects in the inactive executor.
The resolution is to load the installedProjects when executor becomes active. An inactive executor doesn't use installedProjects for any reason.
|
7/10/2018 3:28:16 PM
(#1847)
|
7/10/2018 3:09:24 PM
will pave the way for adding more python based tools.
Future work:
* Run the python tests automatically as other unit tests.
* Separate test code and production code into separate directories.
|
7/10/2018 2:55:17 PM
3.50.0
trigger (#1813)
Upon project uploading, azkaban should be able to schedule the associated flow trigger. This is already achieved when project upload is done from ajax API call(#1631). This PR enables scheduling flow trigger when uploading project from restli endpoint.
|
7/10/2018 1:56:37 PM
parameters. (#1825)
|
7/10/2018 1:10:03 PM
what percentage of code has been exercised by test suites. In order to honor the measurement, This PR adds codecov plugin into travis CI. The detailed documentation of this plugin can be found: https://docs.codecov.io/docs
|
7/10/2018 1:19:35 AM
#1842
Issue:
azkaban.utils.ExecutorServiceUtilsTest > force_shutdown_after_timeout STANDARD_ERROR
log4j:WARN No appenders could be found for logger (azkaban.utils.ExecutorServiceUtils).
log4j:WARN Please initialize the log4j system properly.
log4j:WARN See http://logging.apache.org/log4j/1.2/faq.html#noconfig for more info.
azkaban.utils.ExecutorServiceUtilsTest > force_shutdown_after_timeout FAILED
org.junit.ComparisonFailure: expected:<[tru]e> but was:<[fals]e>
at sun.reflect.NativeConstructorAccessorImpl.newInstance0(Native Method)
at sun.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:62)
at sun.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45)
at azkaban.utils.ExecutorServiceUtilsTest.force_shutdown_after_timeout(ExecutorServiceUtilsTest.java:60)
azkaban.utils.ExecutorServiceUtilsTest
It failed at
assertThat(this.isTestThreadInterrupted).isTrue();
It's possible that the task has not been scheduled yet or has not had time to process the interrupt request.
Fix:
Replaced the test with a mock based test strategy.
The test requires a newer version of JDK. JDK 1.8.05 failed.
1.8.133 worked.
See
https://stackoverflow.com/questions/37527038/mockito-object-is-not-an-instance-of-declaring-class/37528479#37528479
for more information.
|
7/9/2018 7:14:20 PM
3.49.0
Add a test to run a v2 flow file with the new DAG engine
This is a step towards building an integration test for the new DAG engine.
Refactored the DagBuilder API to make it easier to use.
Instead of using a DagBuilder class to link nodes, use the name of the
nodes directly.
This facility can also be used in the future to build the tools to run
flows locally for testing purposes.
* Fix a copy and paste error.
Should check non null of the dagProcessor parameter.
|
7/9/2018 7:07:52 PM
improvements needs to be done on project LRU cache feature:
1. The feature needs to be turned off by default.
2. Running disk calculation on our production clusters whose project dir size is 1.8T takes >= 10 mins.
3. Cleaning thread will be hanging likely due to https://stackoverflow.com/questions/13008526/runtime-getruntime-execcmd-hanging.
4. An alternative design is considered:
Create a file in each project directory and write the size of the project to the file when the project is created. The project files are not supposed to change after creation. Touch this file each time the project is used. This way, we can have a more efficient LRU algorithm based on last access time, not creation time.
Maintain the total size of the project cache in memory to avoid the overhead of re-calculating it. The size shouldn't change too often. This way we can afford to run the check more frequently.
Revert the change to get the release going forward for now and will make the improvements later.
|
7/9/2018 6:12:59 PM
dependency order. Otherwise it can happen that web server tries to launch an execution, but it fails as executor server hasn't started yet.
Also start ExecutorManager's threads (that periodically call executors to get updates) not sooner than when the web server is started. Those threads used to be started already as soon as Guice injection creates the ExecutorManager instance (which happens before executor server is started – even now that the order is fixed).
This is one step towards having everything run in "multi-executor" mode.
As you may know, currently there are two possible modes, controlled by azkaban.use.multiple.executors=true|false:
multi-executor mode, based on the executors DB table
single executor mode, based on fixed props executor.host & executor.port
That adds extra complexity and different branches of code are executed depending on the mode. However, everything can just as well be run as "multi-executor" (but just one executor if that's what you want, and naturally so in case of solo-server). It makes it also easier to reach a better test coverage when testing is always done in the same "mode" that's used in production configurations.
I also noticed that ExecutorManager loads executors from the DB at the time of Guice injection (before executor server is started). This may need to be changed as well to properly support multi-executor mode on a single server.
|
