Hey Dong,
So the main concern with the above approach is that, if for any reason the
> index files of inactive segment is deleted or corrupted, the broker will
> halt if there is only one log directory. This is different from the
> existing behavior where the broker will rebuild the index for this inactive
> segment before it can accept any request from consumer. Though we don't
> have provide guarantee for segments already flushed to disk, this still
> seems like a change in behavior for user. Maybe we don't have to worry
> about this if we decide it is very rare, e.g. it happens only when there is
> disk error or when there is human error.
I think we should probably still handle the case when an index file is
missing during startup? But considering how weak the sanity check is, it
seems fine to skip it. Also, could we just make this change without a KIP?
Adding a config to enable a wimpy sanity check seems unnecessary.
One scenario that does come up with users is actual segment corruption,
which is only detected by consumers that are validating CRCs. To fix it, we
have to manually delete the segments and force re-replication. It would be
helpful to have a config to enable deep checking on startup for particular
topics or partitions. This could also just be a separate tool though
("kafka-fsck" or something).
Thinking longer term, I think we need a more systematic approach to dealing
with corruption, not just in index files, but in the segments as well. It
might be nice, for example, if the consumer had a way to hint the broker
that a particular offset is corrupt. The leader might then demote itself,
for example, and try to repair. Lots to think through though.
-Jason
On Wed, Jun 27, 2018 at 12:29 AM, Dong Lin <lindon...@gmail.com> wrote:
> So the main concern with the above approach is that, if for any reason the
> index files of inactive segment is deleted or corrupted, the broker will
> halt if there is only one log directory. This is different from the
> existing behavior where the broker will rebuild the index for this inactive
> segment before it can accept any request from consumer. Though we don't
> have provide guarantee for segments already flushed to disk, this still
> seems like a change in behavior for user. Maybe we don't have to worry
> about this if we decide it is very rare, e.g. it happens only when there is
> disk error or when there is human error.
>
>
>
> On Wed, Jun 27, 2018 at 12:04 AM, Dong Lin <lindon...@gmail.com> wrote:
>
> > Hey Jason,
> >
> > Thanks for the comment!
> >
> > Your comment reminded me to read through Jay's comments and my reply
> > again. It seems that I probably have not captured idea of Jay's comment
> > that says sanity check is not part of any formal guarantee we provide. I
> > probably should have thought about this comment more. Let me reply to
> both
> > yours and Jay's comment and see if I can understand you better.
> >
> > Here are some clarifications:
> > - KIP does not intend to optimize recovery. It aims to optimize the the
> > sanity check when there is clean shutdown.
> > - Sanity check only read the last entry of the index rather than the full
> > index
> > - We have already done data driven investigation though it is not done
> > using hprof or strace. The resulting rolling bounce time is acceptable
> now.
> > If it appears to be an issue e.g. after more data then we may need to
> > revisit this with more data driven investigation
> >
> > I agree with the following comments:
> > - We should optimize the default behavior instead of adding a new config.
> > - sanity check of the segments before recovery offset is not part of any
> > formal guarantee and thus we probably can just skip it.
> >
> > So we are all leaning towards skipping the sanity check of all segments
> > before the recovery offset. This solution would be pretty straightforward
> > to understand and implement. And I am sure it will give us all the
> benefits
> > that this KIP intends to achieve. Here is only one question to double
> check:
> >
> > If consumer fetches from an inactive segment, broker will just use the
> > index of that inactive segment. If anything goes wrong, e.g. the index
> file
> > is corrupted or the index file does not exist, then the broker will just
> > consider it as IOException, mark the disk and the partitions on the disk
> > offline and respond KafkaStorageException to consumer. Does this sound
> OK?
> > One alternative solution is to let broker rebuild index. But this
> > alternative solution is inconsistent with the idea that "sanity check is
> not
> > part of any formal guarantee" and it may tie up all request handler
> > thread for rebuilding the indexed.
> >
> >
> > If this solution sounds right, I will update the KIP accordingly.
> >
> > Thanks,
> > Dong
> >
> > On Tue, Jun 26, 2018 at 3:23 PM, Jason Gustafson <ja...@confluent.io>
> > wrote:
> >
> >> Hey Dong,
> >>
> >> Sorry for being slow to catch up to this.
> >>
> >> I think the benefit of the sanity check seems a little dubious in the
> >> first
> >> place. We detect garbage at the end of the index file, but that's about
> >> it.
> >> Is there any reason to think that corruption is more likely to occur
> there
> >> or any other reason to think this check is still beneficial for flushed
> >> data? I assume we did the check because we presumed it was cheap, but
> >> perhaps the cost is adding up as the number of partitions grows. How
> much
> >> does startup time improve if we skip the sanity check for data earlier
> >> than
> >> the recovery point? Does the lazy loading itself give some additional
> >> benefit beyond skipping the sanity check? As Jay mentions above, the
> >> sanity
> >> checks seem strictly speaking optional. We don't bother checking the
> >> segments themselves for example.
> >>
> >> Thanks,
> >> Jason
> >>
> >>
> >>
> >>
> >> It probably still makes sense for segments beyond the recovery point
> >>
> >> On Wed, Mar 21, 2018 at 9:59 PM, Dong Lin <lindon...@gmail.com> wrote:
> >>
> >> > Hey Jay,
> >> >
> >> > Yeah our existing sanity check only read the last entry in the index
> >> files.
> >> > I must have miscommunicated if I previously said it was reading the
> full
> >> > index. Broker appears to be spending a lot of time just to read the
> last
> >> > entry of index files for every log segment. This is probably because
> OS
> >> > will load a chunk of data that is much larger than the entry itself
> from
> >> > disk to page cache. This KIP tries to make this part of operation
> lazy.
> >> I
> >> > guess you are suggesting that we should just make the lazy loading the
> >> > default behavior?
> >> >
> >> > Yes we currently require manual intervention if the log file is
> >> corrupted,
> >> > i.e. if two messages with the same offset are appended to the disk
> >> > (KAFKA-6488). The sanity check on broker startup is a bit different
> >> since
> >> > it deals with the corruption of index files (e.g. offset index, time
> >> index
> >> > and snapshot files) instead of the log data. In this case if index
> files
> >> > are corrupted broker will automatically recover it by rebuilding the
> >> index
> >> > files using data in the log files, without requiring manual
> >> intervention.
> >> > Thus the design question is whether this should be done before broker
> >> can
> >> > become leader for any partitions -- there is tradeoff between broker
> >> > startup time and risk of delaying user requests if broker need to
> >> rebuild
> >> > index files when it is already leader. I prefer lazy loading to reduce
> >> > broker startup time. Not sure what are the feedback from the community
> >> on
> >> > this issue.
> >> >
> >> > Thanks,
> >> > Dong
> >> >
> >> >
> >> > On Wed, Mar 21, 2018 at 7:36 AM, Jay Kreps <j...@confluent.io> wrote:
> >> >
> >> > > Hey Dong,
> >> > >
> >> > > Makes total sense. What I'm saying is I don't think that the sanity
> >> check
> >> > > is part of any formal guarantee we provide. It is true that
> >> corruption of
> >> > > data flushed to disk will be a potential problem, but I don't think
> >> the
> >> > > sanity check solves that it just has a couple heuristics to help
> >> detect
> >> > > certain possible instances of it, right? In general I think our
> >> > assumption
> >> > > has been that flushed data doesn't disappear or get corrupted and if
> >> it
> >> > > does you need to manually intervene. I don't think people want to
> >> > configure
> >> > > things at this level so what I was suggesting was understanding why
> >> the
> >> > > sanity check is slow and trying to avoid that rather than making it
> >> > > configurable. I think you mentioned it was reading the full index
> into
> >> > > memory. Based on the performance you describe this could be true,
> but
> >> it
> >> > > definitely should not be reading anything but the last entry in the
> >> index
> >> > > so that would be a bug. That read also happens in sanityCheck() only
> >> in
> >> > the
> >> > > time-based index right? In the offset index we do the same read but
> it
> >> > > happens in initialization. If that read is the slow thing it might
> >> make
> >> > > sense to try to remove it or make it lazy in both cases. If it is
> some
> >> > > other part of the code then (e.g. the size check) then that may be
> >> able
> >> > to
> >> > > be avoided entirely (I think by the time we sanity check we already
> >> know
> >> > > the file size from the mapping...). That was what I meant by doing
> >> some
> >> > > data driven analysis. Maybe a quick run with hprof would help
> >> determine
> >> > the
> >> > > root cause of why sanityCheck is slow?
> >> > >
> >> > > -Jay
> >> > >
> >> > > On Tue, Mar 20, 2018 at 12:13 AM Dong Lin <lindon...@gmail.com>
> >> wrote:
> >> > >
> >> > > > Hey Jay,
> >> > > >
> >> > > > Thanks for your comments!
> >> > > >
> >> > > > Yeah recovery is different from the sanity check. They are
> >> correlated
> >> > in
> >> > > > the sense that there may still be corrupted index files even after
> >> > clean
> >> > > > broker shutdown. And in that case if we delay the sanity check
> then
> >> we
> >> > > may
> >> > > > delay the log recovery. The main goal of this KIP is to optimize
> the
> >> > > sanity
> >> > > > check related work so that it does not delay the broker startup
> >> much.
> >> > > >
> >> > > > The KIP mentioned that the sanity check is done using log recovery
> >> > > > background thread. The name "recovery" is mentioned mainly because
> >> the
> >> > > > background thread number is determined using the existing
> >> > > > config num.recovery.threads.per.data.dir. I have updated the KIP
> to
> >> > make
> >> > > > this less confusing.
> >> > > >
> >> > > > It makes a ton of sense to optimize the broker startup time in a
> >> data
> >> > > > driven fashion. The currently optimize is done kind of in this
> >> fashion.
> >> > > The
> >> > > > broker log shows that LogManager.loadLogs() takes a long time in
> >> large
> >> > > > clusters. Then I started broker with cold cache and repeatedly get
> >> > thread
> >> > > > dump to see what are broker threads are doing during
> >> > > LogManager.loadLogs().
> >> > > > Most of the threads are working on sanityCheck() and this
> motivates
> >> the
> >> > > > change in this KIP. Previously broker shutdown time was
> investigated
> >> > in a
> >> > > > similar data driven fashion and optimized with KAFKA-6172 and
> >> > KAFKA-6175.
> >> > > > It seems that the current KIP can reduces the rolling bounce time
> >> of a
> >> > > > large cluster by 50% -- there may be room for further improvement
> >> but
> >> > > maybe
> >> > > > those do not require as big a change (with the caveat described in
> >> the
> >> > > KIP)
> >> > > > as suggested in this KIP.
> >> > > >
> >> > > > It is not clear whether it is safe to just read the latest segment
> >> > > without
> >> > > > sanity checking all previous inactive segment of a given partition
> >> if
> >> > > > transaction is used. Otherwise we probably want to always skip the
> >> > sanity
> >> > > > check of inactive segments without introducing a new config. Maybe
> >> the
> >> > > > developers familiar with the transaction can comment on that?
> >> > > >
> >> > > > Thanks,
> >> > > > Dong
> >> > > >
> >> > > >
> >> > > > On Mon, Mar 19, 2018 at 7:21 PM, Jay Kreps <j...@confluent.io>
> >> wrote:
> >> > > >
> >> > > > > Optimizing startup seems really valuable but I'm a little
> >> confused by
> >> > > > this.
> >> > > > >
> >> > > > > There are two different things:
> >> > > > > 1. Recovery
> >> > > > > 2. Sanity check
> >> > > > >
> >> > > > > The terminology we're using is a bit mixed here.
> >> > > > >
> >> > > > > Recovery means checksumming the log segments and rebuilding the
> >> index
> >> > > on
> >> > > > a
> >> > > > > hard crash. This only happens on unflushed segments, which is
> >> > generally
> >> > > > > just the last segment. Recovery is essential for the correctness
> >> > > > guarantees
> >> > > > > of the log and you shouldn't disable it. It only happens on hard
> >> > crash
> >> > > > and
> >> > > > > is not a factor in graceful restart. We can likely optimize it
> but
> >> > that
> >> > > > > would make most sense to do in a data driven fashion off some
> >> > > profiling.
> >> > > > >
> >> > > > > However there is also a ton of disk activity that happens during
> >> > > > > initialization (lots of checks on the file size, absolute path,
> >> > etc). I
> >> > > > > think these have crept in over time with people not really
> >> realizing
> >> > > this
> >> > > > > code is perf sensitive and java hiding a lot of what is and
> isn't
> >> a
> >> > > file
> >> > > > > operation. One part of this is the sanityCheck() call for the
> two
> >> > > > indexes.
> >> > > > > I don't think this call reads the full index, just the last
> entry
> >> in
> >> > > the
> >> > > > > index, right?. There should be no need to read the full index
> >> except
> >> > > > during
> >> > > > > recovery (and then only for the segments being recovered). I
> >> think it
> >> > > > would
> >> > > > > make a ton of sense to optimize this but I don't think that
> >> > > optimization
> >> > > > > needs to be configurable as this is just a helpful sanity check
> to
> >> > > detect
> >> > > > > common non-sensical things in the index files, but it isn't part
> >> of
> >> > the
> >> > > > > core guarantees, in general you aren't supposed to lose
> committed
> >> > data
> >> > > > from
> >> > > > > disk, and if you do we may be able to fail faster but we
> >> > fundamentally
> >> > > > > can't really help you. Again I think this would make the most
> >> sense
> >> > to
> >> > > do
> >> > > > > in a data driven way, if you look at that code I think it is
> doing
> >> > > crazy
> >> > > > > amounts of file operations (e.g. getAbsolutePath, file sizes,
> >> etc). I
> >> > > > think
> >> > > > > it'd make most sense to profile startup with a cold cash on a
> >> large
> >> > log
> >> > > > > directory and do the same with an strace to see how many
> redundant
> >> > > system
> >> > > > > calls we do per segment and what is costing us and then cut some
> >> of
> >> > > this
> >> > > > > out. I suspect we could speed up our startup time quite a lot if
> >> we
> >> > did
> >> > > > > that.
> >> > > > >
> >> > > > > For example we have a bunch of calls like this:
> >> > > > >
> >> > > > > require(len % entrySize == 0,
> >> > > > >
> >> > > > > "Index file " + file.getAbsolutePath + " is corrupt,
> >> > found
> >> > > "
> >> > > > +
> >> > > > > len +
> >> > > > >
> >> > > > > " bytes which is not positive or not a multiple of
> >> 8.")
> >> > > > > I'm pretty such file.getAbsolutePath is a system call and I
> assume
> >> > that
> >> > > > > happens whether or not you fail the in-memory check?
> >> > > > >
> >> > > > > -Jay
> >> > > > >
> >> > > > >
> >> > > > > On Sun, Feb 25, 2018 at 10:27 PM, Dong Lin <lindon...@gmail.com
> >
> >> > > wrote:
> >> > > > >
> >> > > > > > Hi all,
> >> > > > > >
> >> > > > > > I have created KIP-263: Allow broker to skip sanity check of
> >> > inactive
> >> > > > > > segments on broker startup. See
> >> > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-
> >> > > > > > 263%3A+Allow+broker+to+skip+sanity+check+of+inactive+
> >> > > > > > segments+on+broker+startup
> >> > > > > > .
> >> > > > > >
> >> > > > > > This KIP provides a way to significantly reduce time to
> rolling
> >> > > bounce
> >> > > > a
> >> > > > > > Kafka cluster.
> >> > > > > >
> >> > > > > > Comments are welcome!
> >> > > > > >
> >> > > > > > Thanks,
> >> > > > > > Dong
> >> > > > > >
> >> > > > >
> >> > > >
> >> > >
> >> >
> >>
> >
> >
>
