I'm thinking that anyone who actually uses compaction has non-standard configuration (at the very least, they had to enable the cleaner, and probably few other configurations too... Compaction is a bit fiddly from what I've seen).
So, I'm in favor of minimal default buffer just for offsets and copycat configuration. If you need compaction elsewhere, feel free to resize. Gwen On Tue, Dec 15, 2015 at 2:29 PM, Grant Henke <ghe...@cloudera.com> wrote: > Following up based on some digging. There are some upper and lower bounds > on the buffer size: > > log.cleaner.dedupe.buffer.size has a: > > - Minimum of 1 MiB per cleaner thread > - > > https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/server/KafkaConfig.scala#L950 > - Maximum of 2 GiB per cleaner thread > - > > https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/log/LogCleaner.scala#L183 > > Then entry size is 24 bytes (the size of an MD5 hash (16 Bytes) + the size > of an offset (8 bytes)). > Note: The hash algorithm is technically interchangeable, but not exposed > via configuration. > > I would like to enable the log cleaner by default given that the new > consumer depends on it. The main concern I have with enabling the log > cleaner by default, and not changing the default size of the dedupe buffer > is the impact it would have on small POC and test deployments that have > small heaps. When moving to the new version, many would just fail with an > OufOfMemoryError. > > We could scale the size of the dedupe buffer down to some percentage of the > maximum memory available not exceeding the configured > log.cleaner.dedupe.buffer.size, and warn if it is less than the configured > value. But I am not sure if that is the best way to handle that either. > > > On Tue, Dec 15, 2015 at 11:23 AM, Jay Kreps <j...@confluent.io> wrote: > > > The buffer determines the maximum number of unique keys in the new > > writes that can be processed in one cleaning. Each key requires 24 > > bytes of space iirc, so 500 MB = ~21,845,333 unique keys (this is > > actually adjusted for some load factor and divided by the number of > > cleaner threads). If it is too small, you will need multiple cleanings > > to compact new data, so sizing too small will tend to lead to lots of > > additional I/O. The tradeoff is that if we size it for just handling > > the offset topic it could be super small (proportional to the number > > of active group-partition combinations), but then people who use log > > compaction will see poor performance. If we size it larger than we > > waste memory. > > > > -Jay > > > > -Jay > > > > On Tue, Dec 15, 2015 at 8:19 AM, Grant Henke <ghe...@cloudera.com> > wrote: > > > Thanks for the background context Jay. > > > > > > Do we have any context on what size is small (but still effect for > small > > > deployments) for the compaction buffer? and what is large? what factors > > > help you choose the correct (or a safe) size? > > > > > > Currently the default "log.cleaner.dedupe.buffer.size" is 500 MiB. If > we > > > are enabling the log cleaner by default, should we adjust that default > > size > > > to be smaller? > > > > > > On a similar note, log.cleaner.delete.retention.ms is currently > > defaulted > > > to 1 day. I am not sure the background here either, but would it make > > sense > > > to default this setting to 7 days to match the default log retention > and > > > ensure no delete messages are missed by consumers? > > > > > > Thanks, > > > Grant > > > > > > On Mon, Dec 14, 2015 at 2:19 PM, Jay Kreps <j...@confluent.io> wrote: > > > > > >> The reason for disabling it by default was (1) general paranoia about > > >> log compaction when we released it, (2) avoid allocating the > > >> compaction buffer. The first concern is now definitely obsolete, but > > >> the second concern is maybe valid. Basically that compaction buffer is > > >> a preallocated chunk of memory used in compaction and is closely tied > > >> to the efficiency of the compaction process (so you want it to be > > >> big). But if you're not using compaction then it is just wasting > > >> memory. I guess since the new consumer requires native offsets > > >> (right?) and native offsets require log compaction, maybe we should > > >> just default it to on... > > >> > > >> -Jay > > >> > > >> On Mon, Dec 14, 2015 at 11:51 AM, Jason Gustafson <ja...@confluent.io > > > > >> wrote: > > >> > That's a good point. It doesn't look like there's any special > handling > > >> for > > >> > the offsets topic, so enabling the cleaner by default makes sense to > > me. > > >> If > > >> > compaction is not enabled, it would grow without bound, so I wonder > > if we > > >> > should even deprecate that setting. Are there any use cases where it > > >> needs > > >> > to be disabled? > > >> > > > >> > -Jason > > >> > > > >> > On Mon, Dec 14, 2015 at 11:31 AM, Gwen Shapira <g...@confluent.io> > > >> wrote: > > >> > > > >> >> This makes sense to me. Copycat also works better if topics are > > >> compacted. > > >> >> > > >> >> Just to clarify: > > >> >> log.cleaner.enable = true just makes the compaction thread run, but > > >> doesn't > > >> >> force compaction on any specific topic. You still need to set > > >> >> delete.policy=compact, and we should not change defaults here. > > >> >> > > >> >> On Mon, Dec 14, 2015 at 10:32 AM, Grant Henke <ghe...@cloudera.com > > > > >> wrote: > > >> >> > > >> >> > Since 0.9.0 the internal "__consumer_offsets" topic is being used > > more > > >> >> > heavily. Because this is a compacted topic does > > "log.cleaner.enable" > > >> need > > >> >> > to be "true" in order for it to be compacted? Or is there special > > >> >> handling > > >> >> > for internal topics? > > >> >> > > > >> >> > If log.cleaner.enable=true is required, should we make it true by > > >> >> default? > > >> >> > Or document that is required for normal operation? > > >> >> > > > >> >> > Thanks, > > >> >> > Grant > > >> >> > -- > > >> >> > Grant Henke > > >> >> > Software Engineer | Cloudera > > >> >> > gr...@cloudera.com | twitter.com/gchenke | > > linkedin.com/in/granthenke > > >> >> > > > >> >> > > >> > > > > > > > > > > > > -- > > > Grant Henke > > > Software Engineer | Cloudera > > > gr...@cloudera.com | twitter.com/gchenke | linkedin.com/in/granthenke > > > > > > -- > Grant Henke > Software Engineer | Cloudera > gr...@cloudera.com | twitter.com/gchenke | linkedin.com/in/granthenke >
