Flink Job cluster in HA mode - recovery vs upgrade
Locked
Flink Job cluster in HA mode - recovery vs upgrade
 
|
Hello,
Let's say I run Flink Job cluster with persistent storage and Zookeeper HA on k8s with single JobManager and use externalized checkpoints. When JM crashes, k8s will restart JM pod, and JM will read JobId and JobGraph from ZK and restore from latest checkpoint.
Now let's say I want to upgrade job binary, I delete deployments, create new deployments referring to newer image, will JM still read JobGraph from ZK or will create new one from new job jar?
Thanks,
Alexey |
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
Hi Alexey, I might be wrong (I don't know this side of Flink very well), but as far as I know JobGraph is never stored in the ZK. It's always recreated from the job's JAR. So you should be able to upgrade the job by replacing the JAR with a newer version, as long as the operator UIDs are the same before and after the upgrade (for operator state to match before and after the upgrade). Best, Piotrek czw., 20 sie 2020 o 06:34 Alexey Trenikhun <[hidden email]> napisał(a):
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
This is incorrect; we do store the
JobGraph in ZooKeeper. If you just delete the deployment the
cluster will recover the previous JobGraph (assuming you aren't
changing the Zookeeper configuration).
If you wish to update the job, then you
should cancel it (along with creating a savepoint), which will
clear the Zookeeper state, and then create a new deployment
On 20/08/2020 15:43, Piotr Nowojski
wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
Thank you for the clarification Chesney and sorry for the incorrect previous answer. Piotrek czw., 20 sie 2020 o 15:59 Chesnay Schepler <[hidden email]> napisał(a):
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
Is it feasible to override ZooKeeperHaServices to recreate JobGraph from jar instead of reading it from ZK state. Any hints? I have feeling that reading JobGraph from jar is more resilient approach, less chances of mistakes during upgrade
Thanks,
Alexey
From: Piotr Nowojski <[hidden email]>
Sent: Thursday, August 20, 2020 7:04 AM To: Chesnay Schepler <[hidden email]> Cc: Alexey Trenikhun <[hidden email]>; Flink User Mail List <[hidden email]> Subject: Re: Flink Job cluster in HA mode - recovery vs upgrade Thank you for the clarification Chesney and sorry for the incorrect previous answer.
Piotrek
czw., 20 sie 2020 o 15:59 Chesnay Schepler <[hidden email]> napisał(a):
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
The HaServices are only being given the
JobGraph, to this is not possible.
Actually I have to correct myself. For
a job cluster the state in HA should be irrelevant when you're
submitting another jar.
Flink has no way of knowing that this
jar is in any way connected to the previous job; they will be
treated as separate things.
However, you will likely end up with
stale data in zookeeper (the jobgraph of the failed job).
On 21/08/2020 17:51, Alexey Trenikhun
wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
Not sure I that I understand your statement about "the HaServices are only being given the JobGraph", seems HighAvailabilityServices#getJobGraphStore provides JobGraphStore, and potentially implementation of JobGraphStore#recoverJobGraph(JobID jobId) for this
store could build new graph for jar rather than read stored graph from ZooKeeper?
Also, if there is single job with same job-id (job cluster), jobgraph of failed job will be over written by new one which will have same job-id?
From: Chesnay Schepler <[hidden email]>
Sent: Friday, August 21, 2020 12:16 PM To: Alexey Trenikhun <[hidden email]>; Piotr Nowojski <[hidden email]> Cc: Flink User Mail List <[hidden email]> Subject: Re: Flink Job cluster in HA mode - recovery vs upgrade The HaServices are only being given the JobGraph, to this is not possible.
Actually I have to correct myself. For a job cluster the state in HA should be irrelevant when you're submitting another jar.
Flink has no way of knowing that this jar is in any way connected to the previous job; they will be treated as separate things.
However, you will likely end up with stale data in zookeeper (the jobgraph of the failed job).
On 21/08/2020 17:51, Alexey Trenikhun wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
If, and only if, the cluster-id and
JobId are identical then the JobGraph will be recovered from
ZooKeeper.
On 22/08/2020 06:12, Alexey Trenikhun
wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
Since it is necessary to use cancel with save point/resume from save point, then it is not possible to use Deployment (otherwise JobManager pod will restart on crash from same save point), so we need to use Job, but in that case
if Job pod is crashed who will start
new instance of Job pod ? Sounds like currently HA with kubernetes is not achievable unless some controller is used to manage JobManager. Am I right?
From: Chesnay Schepler <[hidden email]>
Sent: Saturday, August 22, 2020 12:58 AM To: Alexey Trenikhun <[hidden email]>; Piotr Nowojski <[hidden email]> Cc: Flink User Mail List <[hidden email]> Subject: Re: Flink Job cluster in HA mode - recovery vs upgrade If, and only if, the cluster-id and JobId are identical then the JobGraph will be recovered from ZooKeeper.
On 22/08/2020 06:12, Alexey Trenikhun wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
A job cluster is submitted as a job, not a deployment. The built-in Job controller of Kubernetes ensures that this job
finishes successfully, and if required starts new pods.
On 23/08/2020 06:43, Alexey Trenikhun
wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
Let’s say job cluster was submitted as job from save point sp1, so
spec includes “-s sp1”, job run for days, takin externalized checkpoints every 5 minute, then suddenly pod failed, Kubernetes job controller restarts job pod using original job spec, which has “-s sp1”, so Flink job will
start from sp1 rather than from latest externalized checkpoint. Is my understanding correct?
From: Chesnay Schepler <[hidden email]>
Sent: Sunday, August 23, 2020 1:46:45 AM To: Alexey Trenikhun <[hidden email]>; Piotr Nowojski <[hidden email]> Cc: Flink User Mail List <[hidden email]> Subject: Re: Flink Job cluster in HA mode - recovery vs upgrade A job cluster is submitted as a job, not a deployment. The built-in Job controller of Kubernetes ensures that this job finishes successfully, and if required starts new pods.
On 23/08/2020 06:43, Alexey Trenikhun wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
If HA is enabled the the cluster will
continue from the latest externalized checkpoint.
Without HA it still start from the
savepoint.
On 23/08/2020 16:18, Alexey Trenikhun
wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
Let’s say HA is enabled, so this part works. Now we want to upgrade job jar, we stop job with save point sp2, change manifest to specify “-s sp2” and newer image, and create K8s job again, on start will HAServices still read job
graph from Zookeeper?
From: Chesnay Schepler <[hidden email]>
Sent: Sunday, August 23, 2020 7:25:11 AM To: Alexey Trenikhun <[hidden email]>; Piotr Nowojski <[hidden email]> Cc: Flink User Mail List <[hidden email]> Subject: Re: Flink Job cluster in HA mode - recovery vs upgrade If HA is enabled the the cluster will continue from the latest externalized checkpoint.
Without HA it still start from the savepoint.
On 23/08/2020 16:18, Alexey Trenikhun wrote:
|
Re: Flink Job cluster in HA mode - recovery vs upgrade
|
|
If you do did not specify a different
job or cluster id, then yes it will read the graph from Zookeeper.
Differentiating different submissions
is the very purpose of job ids.
On 23/08/2020 16:38, Alexey Trenikhun
wrote:
|
«
Return to (DEPRECATED) Apache Flink User Mailing List archive.
|
1 view|%1 views
| Free forum by Nabble | Edit this page |