Thanks Liam Clarke-Hutchinson for inputs. Yes, the case that i described is very rare as it is expected *like you mentioned already *the replica failures are transient and eventually caught up once they are back . And the kind of infrastructure we have today, this could eventually be supported by monitoring tools or by using auto heal architectures like kubernetes.
If it ever happens, that is where *as D C suggested *to re assign partitions to start the consuming the messages/ Regards, Nag On Sun, Jun 21, 2020 at 12:11 PM Liam Clarke-Hutchinson < liam.cla...@adscale.co.nz> wrote: > Hi Nag, > > In my experience running Kafka in production for 6 years, so long as the > number of replicas (and the leader is one of those replicas) in the insync > replica set (ISR) is greater than the min.insync.replica setting, the > partition leader will accept writes, and customers can read those writes, > even if the topic is underreplicated. > > Being able to do so is sort of one of the key features of Kafka. > > An out of sync partition replica will catch up when it comes back online, > but until it's a member of the ISR it won't be eligible to become the > partition leader, or for rack aware consumption (a rack aware consumer in > Kafka 2.4.5+ (IIRC) can consume from a replica in the same rack instead of > having to consume from the leader, which can improve latency and in cloud > environments reduce data ingress/egress costs). > > But, I've never encountered a situation where an underreplicated partition > prevents consuming. > > Cheers, > > Liam Clarke-Hutchinson > > On Sun, 21 Jun. 2020, 2:08 pm Nag Y, <andriod.nag.u...@gmail.com> wrote: > > > Thanks D C. Thanks a lot . That is quite a detailed explanation. > > If I understand correctly, ( ignoring the case where producers > > create transactions) - since the replica is down and never comes , the > high > > watermark CANNOT advance and the consumer CAN NOT read the messages which > > were sent after the replica is down as the message is NOT committed - > Hope > > this is correct ? > > > > To address this situation, either we should make sure the replica is up > or > > reduce the replication factor so that the message will be committed and > > consumer can start reading the messages ... > > > > Regards, > > Nag > > > > > > On Sun, Jun 21, 2020 at 3:25 AM D C <dragos.g.cri...@gmail.com> wrote: > > > > > The short answer is : yes, a consumer can only consume messages up to > the > > > High Watermark. > > > > > > The long answer is not exactly, for the following reasons: > > > > > > At the partition level you have 3 major offsets that are important to > the > > > health of the partition and accessibility from the consumer pov: > > > LeO (log end offset) - which represents the highest offset in the > highest > > > segment > > > High Watermark - which represents the latest offset that has been > > > replicated to all the followers > > > LSO (Last stable offset) - which is important when you use producers > that > > > create transactions - which represents the the highest offset that has > > been > > > committed by a transaction and that is allowed to be read with > isolation > > > level = read_commited. > > > > > > The LeO can only be higher or equal to the High Watermark (for obvious > > > reasons) > > > The High Watermark can only be higher or equal to the LSO (the messages > > up > > > to this point may have been committed to all the followers but the > > > transaction isn't yet finished) > > > And coming to your question, in case the transaction hasn't finished, > the > > > LSO may be lower than the High Watermark so if your consumer is > accessing > > > the data in Read_Committed, it won't be able to surpass the LSO. > > > > > > Cheers, > > > D > > > > > > On Sat, Jun 20, 2020 at 9:05 PM Nag Y <andriod.nag.u...@gmail.com> > > wrote: > > > > > > > As I understand it, the consumer can only read "committed" messages - > > > which > > > > I believe, if we look at internals of it, committed messages are > > nothing > > > > but messages which are upto the high watermark. > > > > *The high watermark is the offset of the last message that was > > > successfully > > > > copied to all of the log’s replicas. * > > > > > > > > *Having said that, if one of the replica is down, will high water > mark > > > be* > > > > *advanced?* > > > > > > > > *If replica can't come forever, can we consider this message cant be > > > > consumed by the consumer since it is never committed * > > > > > > > > > >
