On Mon, Jul 23, 2018, at 23:20, James Cheng wrote: > Hi Stanislav! Thanks for this KIP! > > I agree that it would be good if the LogCleaner were more tolerant of > errors. Currently, as you said, once it dies, it stays dead. > > Things are better now than they used to be. We have the metric > kafka.log:type=LogCleanerManager,name=time-since-last-run-ms > which we can use to tell us if the threads are dead. And as of 1.1.0, we > have KIP-226, which allows you to restart the log cleaner thread, > without requiring a broker restart. > https://cwiki.apache.org/confluence/display/KAFKA/KIP-226+-+Dynamic+Broker+Configuration > > <https://cwiki.apache.org/confluence/display/KAFKA/KIP-226+-+Dynamic+Broker+Configuration> > > I've only read about this, I haven't personally tried it.
Thanks for pointing this out, James! Stanislav, we should probably add a sentence or two mentioning the KIP-226 changes somewhere in the KIP. Maybe in the intro section? I think it's clear that requiring the users to manually restart the log cleaner is not a very good solution. But it's good to know that it's a possibility on some older releases. > > Some comments: > * I like the idea of having the log cleaner continue to clean as many > partitions as it can, skipping over the problematic ones if possible. > > * If the log cleaner thread dies, I think it should automatically be > revived. Your KIP attempts to do that by catching exceptions during > execution, but I think we should go all the way and make sure that a new > one gets created, if the thread ever dies. This is inconsistent with the way the rest of Kafka works. We don't automatically re-create other threads in the broker if they terminate. In general, if there is a serious bug in the code, respawning threads is likely to make things worse, by putting you in an infinite loop which consumes resources and fires off continuous log messages. > > * It might be worth trying to re-clean the uncleanable partitions. I've > seen cases where an uncleanable partition later became cleanable. I > unfortunately don't remember how that happened, but I remember being > surprised when I discovered it. It might have been something like a > follower was uncleanable but after a leader election happened, the log > truncated and it was then cleanable again. I'm not sure. James, I disagree. We had this behavior in the Hadoop Distributed File System (HDFS) and it was a constant source of user problems. What would happen is disks would just go bad over time. The DataNode would notice this and take them offline. But then, due to some "optimistic" code, the DataNode would periodically try to re-add them to the system. Then one of two things would happen: the disk would just fail immediately again, or it would appear to work and then fail after a short amount of time. The way the disk failed was normally having an I/O request take a really long time and time out. So a bunch of request handler threads would basically slam into a brick wall when they tried to access the bad disk, slowing the DataNode to a crawl. It was even worse in the second scenario, if the disk appeared to work for a while, but then failed. Any data that had been written on that DataNode to that disk would be lost, and we would need to re-replicate it. Disks aren't biological systems-- they don't heal over time. Once they're bad, they stay bad. The log cleaner needs to be robust against cases where the disk really is failing, and really is returning bad data or timing out. > > * For your metrics, can you spell out the full metric in JMX-style > format, such as: > kafka.log:type=LogCleanerManager,name=uncleanable-partitions-count > value=4 > > * For "uncleanable-partitions": topic-partition names can be very long. > I think the current max size is 210 characters (or maybe 240-ish?). > Having the "uncleanable-partitions" being a list could be very large > metric. Also, having the metric come out as a csv might be difficult to > work with for monitoring systems. If we *did* want the topic names to be > accessible, what do you think of having the > kafka.log:type=LogCleanerManager,topic=topic1,partition=2 > I'm not sure if LogCleanerManager is the right type, but my example was > that the topic and partition can be tags in the metric. That will allow > monitoring systems to more easily slice and dice the metric. I'm not > sure what the attribute for that metric would be. Maybe something like > "uncleaned bytes" for that topic-partition? Or time-since-last-clean? Or > maybe even just "Value=1". I haven't though about this that hard, but do we really need the uncleanable topic names to be accessible through a metric? It seems like the admin should notice that uncleanable partitions are present, and then check the logs? > > * About `max.uncleanable.partitions`, you said that this likely > indicates that the disk is having problems. I'm not sure that is the > case. For the 4 JIRAs that you mentioned about log cleaner problems, all > of them are partition-level scenarios that happened during normal > operation. None of them were indicative of disk problems. I don't think this is a meaningful comparison. In general, we don't accept JIRAs for hard disk problems that happen on a particular cluster. If someone opened a JIRA that said "my hard disk is having problems" we could close that as "not a Kafka bug." This doesn't prove that disk problems don't happen, but just that JIRA isn't the right place for them. I do agree that the log cleaner has had a significant number of logic bugs, and that we need to be careful to limit their impact. That's one reason why I think that a threshold of "number of uncleanable logs" is a good idea, rather than just failing after one IOException. In all the cases I've seen where a user hit a logic bug in the log cleaner, it was just one partition that had the issue. We also should increase test coverage for the log cleaner. > * About marking disks as offline when exceeding a certain threshold, > that actually increases the blast radius of log compaction failures. > Currently, the uncleaned partitions are still readable and writable. > Taking the disks offline would impact availability of the uncleanable > partitions, as well as impact all other partitions that are on the disk. In general, when we encounter I/O errors, we take the disk partition offline. This is spelled out in KIP-112 ( https://cwiki.apache.org/confluence/display/KAFKA/KIP-112%3A+Handle+disk+failure+for+JBOD ) : > - Broker assumes a log directory to be good after it starts, and mark log > directory as > bad once there is IOException when broker attempts to access (i.e. read or > write) the log directory. > - Broker will be offline if all log directories are bad. > - Broker will stop serving replicas in any bad log directory. New replicas > will only be created > on good log directory. The behavior Stanislav is proposing for the log cleaner is actually more optimistic than what we do for regular broker I/O, since we will tolerate multiple IOExceptions, not just one. But it's generally consistent. Ignoring errors is not. In any case, if you want to tolerate an unlimited number of I/O errors, you can just set the threshold to an infinite value (although I think that would be a bad idea). best, Colin > > -James > > > > On Jul 23, 2018, at 5:46 PM, Stanislav Kozlovski <stanis...@confluent.io> > > wrote: > > > > I renamed the KIP and that changed the link. Sorry about that. Here is the > > new link: > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-346+-+Improve+LogCleaner+behavior+on+error > > > > On Mon, Jul 23, 2018 at 5:11 PM Stanislav Kozlovski <stanis...@confluent.io> > > wrote: > > > >> Hey group, > >> > >> I created a new KIP about making log compaction more fault-tolerant. > >> Please give it a look here and please share what you think, especially in > >> regards to the points in the "Needs Discussion" paragraph. > >> > >> KIP: KIP-346 > >> <https://cwiki.apache.org/confluence/display/KAFKA/KIP-346+-+Limit+blast+radius+of+log+compaction+failure> > >> -- > >> Best, > >> Stanislav > >> > > > > > > -- > > Best, > > Stanislav >
