You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@bookkeeper.apache.org by horizonzy <ho...@apache.org> on 2023/07/30 11:00:26 UTC
Data loss problem when DefaultEntryLogger switch to DirectEntryLogger, then switch back to DefaultEntryLogger.
Hi, the followers.
I would need to talk about a serious matter. Especially, if you use
DirectEntryLogger in a production environment.
Switching from DefaultEntryLogger to DirectEntryLogger and switching back
from DirectEntryLogger to DefaultEntryLogger may result in data loss.
Here is the detailed issue https://github.com/apache/bookkeeper/issues/4040
Re: Data loss problem when DefaultEntryLogger switch to DirectEntryLogger, then switch back to DefaultEntryLogger.
Posted by Hang Chen <ch...@apache.org>.
>You should start a new bookie with the new configuration.
If the DirectEntryLogger needs to set up a separate bookie service, it
would be hard for users to switch from the DefaultEntryLogger to
DriectEntryLogger.
>It is not generally possible to guarantee compatibility between different
implementations.
The DirectEntryLogger is designed to be compatible with the
DefaultEntryLogger, and can switch from the DefaultEntryLogger to
DirectEntryLogger smoothly. If we switch from DefaultEntryLogger to
DirectEntryLogger, but can't switch back, it will be hard to apply the
DirectEntryLogger in the existing BookKeeper cluster.
The main reason we can't switch from DirectEntryLogger back to
DefaultEntryLogger is that the EntryLogId generator is different and
the new EntryLogId generator in DirectEntryLogger can cover more cases
than the old one. The drawback of the old EntryLogId generator is as
follows:
- The last entrylogId is stored in the `lastId` file, and if this file
is broken or be tampered, the new generated entrylogId will runs into
unexpected value
- The new entrylogId is generated with +1 logic, and if the entrylogId
reaches Integer.MAX_VALUE, it will be reset to 0 [1]. If the 0.log
still exists (may be impossible in common case), the old 0.log file
will be overridden and cause data loss.
We just need to change the last entryId loader logic in the
DefaultEntryLogger to be compatible with the DirectEntryLogger, which
is valuable for this work.
[1] https://github.com/apache/bookkeeper/blob/3bf08fef12979bf3cff49ee953899f26c6281ad9/bookkeeper-server/src/main/java/org/apache/bookkeeper/bookie/EntryLoggerAllocator.java#L147-L151
Thanks,
Hang
Yan Zhao <ho...@apache.org> 于2023年7月31日周一 17:57写道：
>
> > You should never do this in production.
>
> But the user may not know the limitation, they may believe that the switching is compatible, not introduce some problems.
>
> > BTW do you have a possible solution?
>
> see https://github.com/apache/bookkeeper/pull/4041
Re: Data loss problem when DefaultEntryLogger switch to DirectEntryLogger, then switch back to DefaultEntryLogger.
Posted by Yan Zhao <ho...@apache.org>.
> You should never do this in production.
But the user may not know the limitation, they may believe that the switching is compatible, not introduce some problems.
> BTW do you have a possible solution?
see https://github.com/apache/bookkeeper/pull/4041
Re: Data loss problem when DefaultEntryLogger switch to DirectEntryLogger, then switch back to DefaultEntryLogger.
Posted by Enrico Olivelli <eo...@gmail.com>.
Il giorno lun 31 lug 2023 alle ore 02:40 Yan Zhao
<ho...@apache.org> ha scritto:
>
> I am concerned about a possible scenario where a user tries to use DirectIO, runs it for a period of time, and experiences some problem on the bookie server side, but is unsure of the cause and suspects it may be related to DirectIO. As a result, the user disables DirectIO and restarts. In this case, data may be lost.
You should never do this in production.
BTW do you have a possible solution?
Enrico
>
>
>
> On 2023/07/30 17:40:25 Enrico Olivelli wrote:
> > I think that you should not do this kind of things.
> > You should start a new bookie with the new configuration.
> >
> > It is not generally possible to guarantee compatibility between different
> > implementations.
> >
> > I tend to close the case as 'won't fix'
> >
> >
> > Enrico
> >
> >
> > Il Dom 30 Lug 2023, 13:00 horizonzy <ho...@apache.org> ha scritto:
> >
> > > Hi, the followers.
> > > I would need to talk about a serious matter. Especially, if you use
> > > DirectEntryLogger in a production environment.
> > >
> > > Switching from DefaultEntryLogger to DirectEntryLogger and switching back
> > > from DirectEntryLogger to DefaultEntryLogger may result in data loss.
> > >
> > > Here is the detailed issue
> > > https://github.com/apache/bookkeeper/issues/4040
> > >
> >
Re: Data loss problem when DefaultEntryLogger switch to DirectEntryLogger, then switch back to DefaultEntryLogger.
Posted by Yan Zhao <ho...@apache.org>.
I am concerned about a possible scenario where a user tries to use DirectIO, runs it for a period of time, and experiences some problem on the bookie server side, but is unsure of the cause and suspects it may be related to DirectIO. As a result, the user disables DirectIO and restarts. In this case, data may be lost.
On 2023/07/30 17:40:25 Enrico Olivelli wrote:
> I think that you should not do this kind of things.
> You should start a new bookie with the new configuration.
>
> It is not generally possible to guarantee compatibility between different
> implementations.
>
> I tend to close the case as 'won't fix'
>
>
> Enrico
>
>
> Il Dom 30 Lug 2023, 13:00 horizonzy <ho...@apache.org> ha scritto:
>
> > Hi, the followers.
> > I would need to talk about a serious matter. Especially, if you use
> > DirectEntryLogger in a production environment.
> >
> > Switching from DefaultEntryLogger to DirectEntryLogger and switching back
> > from DirectEntryLogger to DefaultEntryLogger may result in data loss.
> >
> > Here is the detailed issue
> > https://github.com/apache/bookkeeper/issues/4040
> >
>
Re: Data loss problem when DefaultEntryLogger switch to DirectEntryLogger, then switch back to DefaultEntryLogger.
Posted by Enrico Olivelli <eo...@gmail.com>.
I think that you should not do this kind of things.
You should start a new bookie with the new configuration.
It is not generally possible to guarantee compatibility between different
implementations.
I tend to close the case as 'won't fix'
Enrico
Il Dom 30 Lug 2023, 13:00 horizonzy <ho...@apache.org> ha scritto:
> Hi, the followers.
> I would need to talk about a serious matter. Especially, if you use
> DirectEntryLogger in a production environment.
>
> Switching from DefaultEntryLogger to DirectEntryLogger and switching back
> from DirectEntryLogger to DefaultEntryLogger may result in data loss.
>
> Here is the detailed issue
> https://github.com/apache/bookkeeper/issues/4040
>