[
https://issues.apache.org/jira/browse/HUDI-2477?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
sivabalan narayanan updated HUDI-2477:
--------------------------------------
Description:
restore triggers rollback of N commits and then finally will commit the
restore. None of rollbacks will be published to timeline.
But after we have added the rollback.requested instant, restore is breaking w/
metadata enabled.
Here is what is happening:
Restore
schedule rollback for all of N commits. this will produce
rollback.requested instants to timeline. Remember we can't skip this
publishing, bcoz, rollback action executor depends on this.
trigger rollback action executor. which will execute the rollback. but this
time we may not publish the rollbacks. and so there won't be a rollback
completed instant.
and now to finalize restore, we apply the changes to metadata table before we
can commit the restore to datatable. Here is where the issue is. We do check if
bootstrapping is required. chances that last synced instant to metadata table
is not active anymore in data table and so it triggers a bootstrap. but we do
allow bootstrap only if there are no pending operations in datatable. But all
rollbacks are surfacing as pending operations and hence we fail here.
This could also be an issue when we try to play with bootstrap in original
dataset.
bootstrap. and for some reason you want to rollback bootstrap. this might end
up in this state too.
to illustrate clearly.
bootstrap
also apply changes to metadata. there is only one commit.
rollback bootstrap
this is a restore operation. so, we first do a rollback which will create
rollback.requested instant.
and to finalize restore, we try to apply the restore to
metadata.
this goes into bootstrap code path. last synced instant is
not found in datatimeline. we assume its archived and so trigger a rebootstrap
and so delete the metadata table.
and then try to do the actual bootstrap. but since there is
a pending operation in datatimeline (rollback.requested), we will not do any
bootstrap only. and so the state remains. i.e. metadata table is deleted. and
the actually applying restore commit will fail.
We also need to think even if metadata is enabled, should we leave the rollback
instant in timeline itself.or should we clean it up after committing restore to
timeline.
was:
restore triggers rollback of N commits and then finally will commit the
restore. None of rollbacks will be published to timeline.
But after we have added the rollback.requested instant, restore is breaking w/
metadata enabled.
Here is what is happening:
Restore
schedule rollback for all of N commits. this will produce
rollback.requested instants to timeline. Remember we can't skip this
publishing, bcoz, rollback action executor depends on this.
trigger rollback action executor. which will execute the rollback. but this
time we may not publish the rollbacks. and so there won't be a rollback
completed instant.
and now to finalize restore, we apply the changes to metadata table before we
can commit the restore to datatable. Here is where the issue is. We do check if
bootstrapping is required. chances that last synced instant to metadata table
is not active anymore in data table and so it triggers a bootstrap. but we do
allow bootstrap only if there are no pending operations in datatable. But all
rollbacks are surfacing as pending operations and hence we fail here.
This could also be an issue when we try to play with bootstrap in original
dataset.
bootstrap. and for some reason you want to rollback bootstrap. this might end
up in this state too.
We also need to think even if metadata is enabled, should we leave the rollback
instant in timeline itself.or should we clean it up after committing restore to
timeline.
> Restore fails after adding rollback plan and rollback.requested instant w/
> metadata enabled
> -------------------------------------------------------------------------------------------
>
> Key: HUDI-2477
> URL: https://issues.apache.org/jira/browse/HUDI-2477
> Project: Apache Hudi
> Issue Type: Sub-task
> Components: Writer Core
> Reporter: sivabalan narayanan
> Assignee: sivabalan narayanan
> Priority: Major
> Fix For: 0.10.0
>
>
> restore triggers rollback of N commits and then finally will commit the
> restore. None of rollbacks will be published to timeline.
> But after we have added the rollback.requested instant, restore is breaking
> w/ metadata enabled.
> Here is what is happening:
> Restore
> schedule rollback for all of N commits. this will produce
> rollback.requested instants to timeline. Remember we can't skip this
> publishing, bcoz, rollback action executor depends on this.
> trigger rollback action executor. which will execute the rollback. but
> this time we may not publish the rollbacks. and so there won't be a rollback
> completed instant.
> and now to finalize restore, we apply the changes to metadata table before we
> can commit the restore to datatable. Here is where the issue is. We do check
> if bootstrapping is required. chances that last synced instant to metadata
> table is not active anymore in data table and so it triggers a bootstrap. but
> we do allow bootstrap only if there are no pending operations in datatable.
> But all rollbacks are surfacing as pending operations and hence we fail here.
>
> This could also be an issue when we try to play with bootstrap in original
> dataset.
> bootstrap. and for some reason you want to rollback bootstrap. this might end
> up in this state too.
> to illustrate clearly.
> bootstrap
> also apply changes to metadata. there is only one commit.
> rollback bootstrap
> this is a restore operation. so, we first do a rollback which will create
> rollback.requested instant.
> and to finalize restore, we try to apply the restore to
> metadata.
> this goes into bootstrap code path. last synced instant
> is not found in datatimeline. we assume its archived and so trigger a
> rebootstrap and so delete the metadata table.
> and then try to do the actual bootstrap. but since there
> is a pending operation in datatimeline (rollback.requested), we will not do
> any bootstrap only. and so the state remains. i.e. metadata table is deleted.
> and the actually applying restore commit will fail.
>
>
> We also need to think even if metadata is enabled, should we leave the
> rollback instant in timeline itself.or should we clean it up after committing
> restore to timeline.
>
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
