[
https://issues.apache.org/jira/browse/SQOOP-1510?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14191972#comment-14191972
]
Jarek Jarcec Cecho commented on SQOOP-1510:
-------------------------------------------
I think that [~gwenshap] has a good points, so +1 on them. In addition here are
answers from my side:
{quote}
What does SUBMIT and START mean in our cases, what should be the value for the
2 enums we have.
the one in the client that has 3 values ( SUBMITTED, UPDATED, FINISHED) and the
one in the server - SubmissionStatus ( around 8 different values)
{quote}
In the utmost details:
* Submit - refers to state when we've created and submitted job request (using
MR terminology) to the Hadoop cluster.
* Start - refers to state when the job actually started. Usually those two are
almost synonyms, however on busy cluster there can be delay between them.
The
[SubmissionStatus|https://github.com/apache/sqoop/blob/sqoop2/common/src/main/java/org/apache/sqoop/submission/SubmissionStatus.java#L23]
is enum covering all the options that are possible to properly describe what
is happening. We do have helper methods that can be used to get some sense into
the values. For example
[isRunning|https://github.com/apache/sqoop/blob/sqoop2/common/src/main/java/org/apache/sqoop/submission/SubmissionStatus.java#L72]
or
[isFailure|https://github.com/apache/sqoop/blob/sqoop2/common/src/main/java/org/apache/sqoop/submission/SubmissionStatus.java#L76].
The enum in
[SqoopClient|https://github.com/apache/sqoop/blob/sqoop2/client/src/main/java/org/apache/sqoop/client/SqoopClient.java#L84]
is private to the class and is used only to make the internal code (in this
class) easier.
Please do note that we are intentionally not exposing all the technical details
to the end user. For example the difference between "Submit" and "Start" is not
a first class citizen in the shell as the common user don't need to know the
difference whereas we are keeping the difference on the API level.
{quote}
Second, I have explained why we have v1/connector and v1/connectors , it was in
the PDF and design proposed 1 month back. It is common practice to use plural
for the collection resource, I have never seen so far anyone recommend using
ALL for every entity we have, whats the point behind it?
{quote}
I don't recall technical details behind that so I assume that it was done for
simplicity. I don't mind to have both "connector" and "connectors" end point,
even though that I don't see that in practice. In my Ruby on Rails years I've
always had one end point for one entity that was able to serve both 1 entity
and multiple entities.
{quote}
Third, please read the submit method in JobManager. Why is the destry method of
the conenctor called in failure case ? what happens int he case of Kite
connector when there is merge logic n the destoyer.
{quote}
We already called initializer that can have side effects such as new temporary
table(s) on database side, temporary directories on HDFS, ... all those should
be cleaned up and hence we are calling the destroyer with
[parameter|https://github.com/apache/sqoop/blob/sqoop2/common/src/main/java/org/apache/sqoop/job/etl/DestroyerContext.java#L45]
telling it that the job has failed. Kite connector shouldn't have any trouble
with this because it already have to solve the situation when job fails - the
specific here is that the job hasn't even started and hence there is no
OutputCommitter from which it can be called. I'm assuming that on failure there
wont' be any merge logic as the job has failed.
> Refactor JobRequestHandler for submit/abort job and SubmissionHandler for get
> operation only
> --------------------------------------------------------------------------------------------
>
> Key: SQOOP-1510
> URL: https://issues.apache.org/jira/browse/SQOOP-1510
> Project: Sqoop
> Issue Type: Sub-task
> Reporter: Veena Basavaraj
> Assignee: Veena Basavaraj
>
> PROPOSAL JIST:
> 1. Moves the actions of job submit/ status of current running job and
> stopping currently running job to the JOB resource . A submit action creates
> a submission record. To be consistent with the v1/connectors and v1/links we
> also have v1/jobs to display all jobs abd v1/jobs? cname by connector
> Also proposing ( note this can be another RB if it is hard to review the
> current one )
> START-> renamed to SUBMIT
> STOP -> renamed to abort
> 2. Submission is a read only resource that can give all the submissions or
> submissions per job. Since it is collection resource it is uses
> v1/submissions to be consistent with v1/connectors. v1/links and v1/jobs.
> 3. Changes to the client/ shell to reflect the 1 and 2
> PROPOSAL DETAILS
> Adding more details to the ticket to explain how the Job/ Submission are
> structured in the code.
> In sqoop we will create a job giving it the FROM and the TO link ids
> Something like this ...
> {code}
> create job -fromLink 1 -toLink 2
> {code}
> A job in SQOOP is a representation of all the data required for the JOB we
> will submit to the execution engine. hence the job in sqoop will hold the
> FROM link and FROM connector and TO link and its corresponding TO connector
> details. It will hold all its corresponding config values to invoke the FROM
> connector code and the TO connector code as part of the job lifecycle ( init,
> partitions, load, extract, destroy) steps.
> So once the job is created, we can perform these 4 actions on it
> 1. disable / re-enable it
> 2. submit the job to the excution engine
> 3. at any point while it is running, abort it ( we can also call this stop if
> we want to)
> All of the below is handled by the
> {code}JobManager{code} is an internal class that receives the requests for
> doing operations on the job resource.
> So what does submit really do?
> 1. create a job request for the execution engine. This is a UBER object that
> holds all the information I just spoke about earlier, like FROM connector and
> TO connector details. It also holds a reference to the submission object.
> Submission object holds the results of the submit action. A new submission
> record is persisted into the SQOOP repository every time we call a job
> submit. This is representated as {code} MSubmission {code} internally.
> {code}
> // Bootstrap job to execute in the configured execution engine
> prepareJob(jobRequest);
> {code}
> 2. We then call the submissionEngine api to submit, that will inturn choose
> the configured execution engine ( such as MR or spark in future) and then
> submit to it.
> {code}
> boolean success = submissionEngine.submit(jobRequest);
> {code}
> {code}
> / If we're in local mode than wait on completion. Local job runner do not
> // seems to be exposing API to get previously submitted job which makes
> // other methods of the submission engine quite useless.
> if(isLocal()) {
> job.waitForCompletion(true);
> } else {
> job.submit();
> }
> {code}
> For consistency we call the submission and execution apis as submit since MR
> engine and spark also use submit
> 3. Once we succeed we persist the record in the Sqoop repo.
> {code}
>
> RepositoryManager.getInstance().getRepository().createSubmission(mJobSubmission);
> {code}
> 4. If execution engine failed, we call some clean up code, but still persist
> the submission record in repo to record the "submit" action invoked from the
> client/ rest api
> {code}
> public MSubmission submit(long jobId, HttpEventContext ctx) {
> MSubmission mJobSubmission = createJobSubmission(ctx, jobId);
> JobRequest jobRequest = createJobRequest(jobId, mJobSubmission);
> // Bootstrap job to execute in the configured execution engine
> prepareJob(jobRequest);
> // Make sure that this job id is not currently running and submit the job
> // only if it's not.
> synchronized (getClass()) {
> MSubmission lastSubmission =
> RepositoryManager.getInstance().getRepository()
> .findLastSubmissionForJob(jobId);
> if (lastSubmission != null && lastSubmission.getStatus().isRunning()) {
> throw new SqoopException(DriverError.DRIVER_0002, "Job with id " +
> jobId);
> }
> // TODO(jarcec): We might need to catch all exceptions here to ensure
> // that Destroyer will be executed in all cases.
>
> // NOTE: the following is a blocking call
> boolean success = submissionEngine.submit(jobRequest);
> if (!success) {
> cleanUpOnJobSubmissionFailure(jobRequest);
> mJobSubmission.setStatus(SubmissionStatus.FAILURE_ON_SUBMIT);
> }
> // persist submisison record to repository and success status
>
> RepositoryManager.getInstance().getRepository().createSubmission(mJobSubmission);
> }
> return mJobSubmission;
> }
> {code}
> Next, What happens if we want to stop/abort the running job.
> 1. The stop action is a nooperation if there is no running job currently. We
> have an api that tells the last job submitted and we use it to check if the a
> requested job is in running state.
> {code}
> // remove the last running job submission
> MSubmission mSubmission = repository.findLastSubmissionForJob(jobId);
> if (mSubmission == null || !mSubmission.getStatus().isRunning()) {
> throw new SqoopException(DriverError.DRIVER_0003, "Job with id " + jobId
> + " is not running hence cannot stop");
> }
>
> {code}
> 2. If a job was running, we call the submission engine api to stop which then
> calls the corresponding execution engine apis "kill Job"
> {code} submissionEngine.stop(mSubmission.getExternalId()); {code} For
> consistency we call all the apis as stop /abort
> {code}RunningJob runningJob = jobClient.getJob(JobID.forName(submissionId));
> if(runningJob == null) {
> return;
> }
> runningJob.killJob(); {code}
> 3. Finally we update the status of the submission in the repo by calling
> {code}updateSubmission{code}
> So in summary the actions on JOB will be reflected on the job resource
> The job will have enable/ disable/ submit / abort(stop) Also get on the
> status and the job details
> Submission is a read only resource which represents the side affect of
> submitting a job to the sqoop execution engine
> While we change the rest apis to support the above, we also have to make the
> corresponsing changes in the sqoop client.
> Here is the details of the sqoop client. {code}SqoopClient{code} will support
> the submit/ abort(stop) a job
> 1. The client has limited visibility into the job submission. After a job
> submit command is issued it can monitor the status of the job that is
> submitted.
> Hence we have this enum JobSubmissionStatus, at any point we will reply back
> with one of the following 3 states. Started means it just got kicked off.
> Updated mean
> {code}
> /**
> * Status flags used when updating the job submission callback status
> */
> private enum JobSubmissionStatus {
> STARTED,
> UPDATED,
> FINISHED
> }
> {code}
> 2. The client will call the {code}SqoopResourceRequest {code}method to make a
> rest call for the "submit" action and get back a submission record. It can
> use this submission record to query the status if need be and get a more
> real-time feedback to client on the status of the job using submisison
> callback logic.
> {code}
> MSubmission submission =
> resourceRequests.submitJob(jobId).getSubmissions().get(0);
> {code}
> {code}
> public MSubmission submitJob(long jobId, SubmissionCallback callback, long
> pollTime)
> throws InterruptedException {
> if(pollTime <= 0) {
> throw new SqoopException(ClientError.CLIENT_0002);
> }
> boolean started = true;
> MSubmission submission =
> resourceRequests.submitJob(jobId).getSubmissions().get(0);
> while(submission.getStatus().isRunning()) {
> if(started) {
> submissionCallback(callback, submission, JobSubmissionStatus.STARTED);
> started = false;
> } else {
> submissionCallback(callback, submission, JobSubmissionStatus.UPDATED);
> }
> Thread.sleep(pollTime);
> submission = getJobStatus(jobId);
> }
> submissionCallback(callback, submission, JobSubmissionStatus.FINISHED);
> return submission;
> }
> {code}
> 3. Similarly for the stop action it calls the JobResoource to make a rest
> call to the server for stopping the job
> {code}
> public MSubmission stopJob(long jid) {
> return resourceRequests.stopJob(jid).getSubmissions().get(0);
> }
> {code}
> Why is the action called SUBMIT/ instead of START?
> SUBMIT is a common term in both MR and SPARK execution engine, Also as a side
> affect of SUBMIT we create a SUBMISSION record and store it in the repository.
> There are tons of places in the code and in the java docs, we actually mean
> submit when we say start in the current code. So why not just call it submit.
> DRIVER_0008("Invalid combination of submission and execution engines"),
> DRIVER_0009("Job has been disabled. Cannot submit this job."),
> DRIVER_0010("Link for this job has been disabled. Cannot submit this job."),
> DRIVER_0011("Connector does not support specified direction. Cannot submit
> this job."),
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
