Hello Xiao, The proposed transactional messaging at Kafka is aimed at improving the at-least-once semantics of delivery to exactly-once, i.e. to avoid duplicates. It is not aimed for grouping fsync of multiple messages into one, as for avoiding data loss it is still dependent on data replication.
As I understands your situation, you are mainly concerned with data center power outage tolerance via disk persistence, which cannot be easily supported via asynchronous time-based fsync. We used to do synchronized fsync before data replication is introduced, and later switched to async fsync for mainly for performance (KAFKA-615 <https://issues.apache.org/jira/browse/KAFKA-615>), since we see the disk IO spikes can some times cause the whole broker to slow down quite a lot due to processor thread pool sharing. On the other hand, even with fsync, data loss may still happen upon power outage depending on your file system journaling settings. Another alternative to achieve data center failures is geo-replication: you can try to put your replicas in different data centers, more specifically: 1) you can use one producer to send normal data with ack mode = 1 for not waiting cross-colo replication latency. 2) at each batch boundary, you can use another producer to send a "commit" message with ack mode = -1, which requires all data before this message to be already replicated, to mimic the "group committing" behavior. This will of course still increase the write latency and require a cross-colo ZK setup. Guozhang On Mon, Mar 9, 2015 at 10:56 PM, Xiao <lixiao1...@gmail.com> wrote: > Hi, Jay, > > Thank you! > > The Kafka document shows “Kafka should run well on any unix system". I > assume it includes the major two Unix versions, IBM AIX and HP-UX. Right? > > 1. Unfortunately, we aims at supporting all the platforms, Linux, Unix, > Windows and especially z/OS. I know z/OS is not easy to support. > > 2. Fsync per message is very expensive and Fsync per batch will break the > transaction atomicity. We are looking for transaction-level fsync, which is > more efficient. Then, our producers can easily combine multiple small > transactions into a single bigger batch transaction. I hope the > implementation of the ongoing Kafka feature “Transactional Messaging in > Kafka” already considered all these issues, although the following link > does not mention it: > https://cwiki.apache.org/confluence/display/KAFKA/Transactional+Messaging+in+Kafka > > 3. After a quick reading of the design of “Transactional Messaging in > Kafka”, I have a doubt if it can scale especially when most transactions > are very short (e.g., containing a single message). > - In many use cases, we do not need global transactions, which > might be too expensive. The partition-specific transaction granularity > might be fine. > - To reduce the overhead of transaction-level fsync + network or > address-space roundtrips, Kafka might need two extra parameters for > batching the small transactions into a single one. The performance benefits > of batching are huge, as shown in the MRI (multi-row insert) feature in > Oracle and DB2 z/OS. > > I believe transactional messaging is a critical feature. The design > document is not very clear. Do you have more materials or links about it? > > Thanks, > > Xiao Li > > > On Mar 7, 2015, at 9:33 AM, Jay Kreps <jay.kr...@gmail.com> wrote: > > > Xiao, > > > > FileChannel.force is fsync on unix. > > > > To force fsync on every message: > > log.flush.interval.messages=1 > > > > You are looking at the time based fsync, which, naturally, as you say, is > > time-based. > > > > -Jay > > > > On Fri, Mar 6, 2015 at 11:35 PM, Xiao <lixiao1...@gmail.com> wrote: > > > >> Hi, Jay, > >> > >> Thank you for your answer. > >> > >> Sorry, I still do not understand your meaning. > >> > >> I guess the two parameters you mentioned are log.flush.interval and > >> log.default.flush.interval.ms. However, these two parameters only > control > >> when Kafka issues a flush (i.e., calling FileChannel.force()). > >> > >> Fsync (fileOutputStream.getFD().sync()) is controlled by another > parameter > >> log.default.flush.scheduler.interval.ms. > >> > >> scheduler.schedule("kafka-recovery-point-checkpoint", > >> checkpointRecoveryPointOffsets, > >> delay = InitialTaskDelayMs, > >> period = flushCheckpointMs, > >> TimeUnit.MILLISECONDS) > >> > >> This thread is only time-controlled. It does not check the number of > >> messages. > >> > >> Thank you, > >> > >> Xiao Li > >> > >> > >> On Mar 5, 2015, at 11:59 AM, Jay Kreps <jay.kr...@gmail.com> wrote: > >> > >>> Hey Xiao, > >>> > >>> That's not quite right. Fsync is controlled by either a time based > >> criteria > >>> (flush every 30 seconds) or a number of messages criteria. So if you > set > >>> the number of messages to 1 the flush is synchronous with the write, > >> which > >>> I think is what you are looking for. > >>> > >>> -Jay > >> > >> > > -- -- Guozhang
