Excellent. Thanks for all the answers so far. So there was another issue I mentioned which we made some progress gaining insight into, namely our metaspace growth when faced with job restarts.
We can easily hit 1Gb metaspace usage within 15 minutes if we restart often. We attempted to troubleshoot this issue by looking at all the classes in metaspace using `jcmd <pid> GC.class_stats`. Here we observed that after every job restart another entry is created for every class in our job. Where the old classes have InstBytes=0. So far so good, but moving to the Total column for these entries show that memory is still being used. Also, adding up all entries in the Total column indeed corresponds to our metaspace usage. So far we could only conclude that our job classes - none of them - were being unloaded. Then we stumbled upon this ticket. Now here are our results running the SocketWindowWordCount jar in a flink 1.8.0 cluster with one taskmanager. We achieve a class count by doing a jcmd 3052 GC.class_stats | grep -i org.apache.flink.streaming.examples.windowing.SessionWindowing | wc -l *First* run: Class Count: 1 Metaspace: 30695K After *800*~ runs: Class Count: 802 Metaspace: 39406K Interesting when we looked a bit later the class count *slowly* went down, slowly, step by step, where just to be sure we used `jcmd <pid> GC.run` to force GC every 30s or so. If I had to guess it took about 20 minutes to go from 800~ to 170~, with metaspace dropping to 35358K. In a sense we've seen this behavior, but with much much larger increases in metaspace usage over far fewer job restarts. I've added this information to https://issues.apache.org/jira/browse/FLINK-11205. That said, I'd really like to confirm the following: - classes should usually only appear once in GC.class_stats output - flink / the jvm has very slow cleanup of the metaspace - something clearly is leaking during restarts On Mon, Jul 29, 2019 at 9:52 AM Yu Li <car...@gmail.com> wrote: > For the memory usage of RocksDB, there's already some discussion in > FLINK-7289 <https://issues.apache.org/jira/browse/FLINK-7289> and a good > suggestion > <https://issues.apache.org/jira/browse/FLINK-7289?focusedCommentId=16874305&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-16874305> > from Mike to use the WriteBufferManager to limit the total memory usage, > FYI. > > We will drive to make the memory management of state backends more "hands > free" in latter release (probably in release 1.10) and please watch the > release plan and/or the weekly community update [1] threads. > > [1] https://s.apache.org/ix7iv > > Best Regards, > Yu > > > On Thu, 25 Jul 2019 at 15:12, Yun Tang <myas...@live.com> wrote: > >> Hi >> >> It's definitely not easy to calculate the accurate memory usage of >> RocksDB, but formula of "block-cache-memory + column-family-number * >> write-buffer-memory * write-buffer-number + index&filter memory" should >> give enough sophisticated hints. >> When talking about the column-family-number, they are equals to the >> number of your states which are the declared state descriptors in one >> operator and potential one window state (if you're using window). >> The default writer-buffer-number is 2 at most for each column family, and >> the default write-buffer-memory size is 4MB. Pay attention that if you ever >> configure the options for RocksDB, these memory usage would differ from >> default values. >> The last part of index&filter memory is not easy to estimate, but from my >> experience this part of memory would not occupy too much only if you have >> many open files. >> >> Last but not least, Flink would enable slot sharing by default, and even >> if you only one slot per taskmanager, there might exists many RocksDB >> within that TM due to many operator with keyed state running. >> >> Apart from the theoretical analysis, you'd better to open RocksDB native >> metrics or track the memory usage of pods through Prometheus with k8s. >> >> Best >> Yun Tang >> ------------------------------ >> *From:* wvl <lee...@gmail.com> >> *Sent:* Thursday, July 25, 2019 17:50 >> *To:* Yang Wang <danrtsey...@gmail.com> >> *Cc:* Yun Tang <myas...@live.com>; Xintong Song <tonysong...@gmail.com>; >> user <user@flink.apache.org> >> *Subject:* Re: Memory constrains running Flink on Kubernetes >> >> Thanks for all the answers so far. >> >> Especially clarifying was that RocksDB memory usage isn't accounted for >> in the flink memory metrics. It's clear that we need to experiment to >> understand it's memory usage and knowing that we should be looking at the >> container memory usage minus all the jvm managed memory, helps. >> >> In mean while, we've set MaxMetaspaceSize to 200M based on our metrics. >> Sadly the resulting OOM does not result a better behaved job, because it >> would seem that the (taskmanager) JVM itself is not restarted - which makes >> sense in a multijob environment. >> So we're looking into ways to simply prevent this metaspace growth (job >> library jars in /lib on TM). >> >> Going back to RocksDB, the given formula "block-cache-memory + >> column-family-number * write-buffer-memory * write-buffer-number + >> index&filter memory." isn't completely clear to me. >> >> Block Cache: "Out of box, RocksDB will use LRU-based block cache >> implementation with 8MB capacity" >> Index & Filter Cache: "By default index and filter blocks are cached >> outside of block cache, and users won't be able to control how much memory >> should be use to cache these blocks, other than setting max_open_files.". >> The default settings doesn't set max_open_files and the rocksdb default >> seems to be 1000 ( >> https://github.com/facebook/rocksdb/blob/master/include/rocksdb/utilities/leveldb_options.h#L89) >> .. not completely sure about this. >> Write Buffer Memory: "The default is 64 MB. You need to budget for 2 x >> your worst case memory use." >> >> May I presume a unique ValueStateDescriptor equals a Column Family? >> If so, say I have 10 of those. >> 8MB + (10 * 64 * 2) + $Index&FilterBlocks >> >> So is that correct and how would one calculate $Index&FilterBlocks? The >> docs suggest a relationship between max_open_files (1000) and the amount >> index/filter of blocks that can be cached, but is this a 1 to 1 >> relationship? Anyway, this concept of blocks is very unclear. >> >> > Have you ever set the memory limit of your taskmanager pod when >> launching it in k8s? >> >> Definitely. We settled on 8GB pods with taskmanager.heap.size: 5000m and >> 1 slot and were looking into downsizing a bit to improve our pod to VM >> ratio. >> >> On Wed, Jul 24, 2019 at 11:07 AM Yang Wang <danrtsey...@gmail.com> wrote: >> >> Hi, >> >> >> The heap in a flink TaskManager k8s pod include the following parts: >> >> - jvm heap, limited by -Xmx >> - jvm non-heap, limited by -XX:MaxMetaspaceSize >> - jvm direct memory, limited by -XX:MaxDirectMemorySize >> - native memory, used by rocksdb, just as Yun Tang said, could be >> limited by rocksdb configurations >> >> >> So if your k8s pod is terminated by OOMKilled, the cause may be the >> non-heap memory or native memory. I suggest you add an environment >> FLINK_ENV_JAVA_OPTS_TM="-XX:MaxMetaspaceSize=512m" in your >> taskmanager.yaml. And then only the native memory could cause OOM. Leave >> enough memory for rocksdb, and then hope your job could run smoothly. >> >> Yun Tang <myas...@live.com> 于2019年7月24日周三 下午3:01写道： >> >> Hi William >> >> Have you ever set the memory limit of your taskmanager pod when launching >> it in k8s? If not, I'm afraid your node might come across node >> out-of-memory [1]. You could increase the limit by analyzing your memory >> usage >> When talking about the memory usage of RocksDB, a rough calculation >> formula could be: block-cache-memory + column-family-number * >> write-buffer-memory * write-buffer-number + index&filter memory. The block >> cache, write buffer memory&number could be mainly configured. And the >> column-family number is decided by the state number within your operator. >> The last part of index&filter memory cannot be measured well only if you >> also cache them in block cache [2] (but this would impact the performance). >> If you want to the memory stats of rocksDB, turn on the native metrics of >> RocksDB [3] is a good choice. >> >> >> [1] >> https://kubernetes.io/docs/tasks/administer-cluster/out-of-resource/#node-oom-behavior >> [2] >> https://github.com/facebook/rocksdb/wiki/Memory-usage-in-RocksDB#indexes-and-filter-blocks >> [3] >> https://ci.apache.org/projects/flink/flink-docs-release-1.8/ops/config.html#rocksdb-native-metrics >> >> Best >> Yun Tang >> ------------------------------ >> *From:* Xintong Song <tonysong...@gmail.com> >> *Sent:* Wednesday, July 24, 2019 11:59 >> *To:* wvl <lee...@gmail.com> >> *Cc:* user <user@flink.apache.org> >> *Subject:* Re: Memory constrains running Flink on Kubernetes >> >> Hi, >> >> Flink acquires these 'Status_JVM_Memory' metrics through the MXBean >> library. According to MXBean document, non-heap is "the Java virtual >> machine manages memory other than the heap (referred as non-heap memory)". >> Not sure whether that is equivalent to the metaspace. If the >> '-XX:MaxMetaspaceSize', it should trigger metaspcae clean up when the limit >> is reached. >> >> As for RocksDB, it mainly uses non-java memory. Heap, non-heap and direct >> memory could be considered as java memory (or at least allocated through >> the java process). That means, RocksDB is actually using the memory that is >> accounted in the total K8s container memory but not accounted in neither of >> java heap / non-heap / direct memory, which in your case the 1GB >> unaccounted. To leave more memory for RocksDB, you need to either configure >> more memory for the K8s containers, or configure less java memory through >> the config option 'taskmanager.heap.size'. >> >> The config option 'taskmanager.heap.size', despite the 'heap' in its key, >> also accounts for network memory (which uses direct buffers). Currently, >> memory configurations in Flink is quite complicated and confusing. The >> community is aware of this, and is planing for an overall improvement. >> >> To my understanding, once you set '-XX:MaxMetaspaceSize', there should be >> limits on heap, non-heap and direct memory in JVM. You should be able to >> find which part that requires memory more than the limit from the java OOM >> error message. If there is no java OOM but a K8s container OOM, then it >> should be non-java memory used by RocksDB. >> >> [1] >> https://docs.oracle.com/javase/8/docs/api/java/lang/management/MemoryMXBean.html >> >> Thank you~ >> >> Xintong Song >> >> >> >> On Tue, Jul 23, 2019 at 8:42 PM wvl <lee...@gmail.com> wrote: >> >> Hi, >> >> We're running a relatively simply Flink application that uses a bunch of >> state in RocksDB on Kubernetes. >> During the course of development and going to production, we found that >> we were often running into memory issues made apparent by Kubernetes >> OOMKilled and Java OOM log events. >> >> In order to tackle these, we're trying to account for all the memory used >> in the container, to allow proper tuning. >> Metric-wise we have: >> - container_memory_working_set_bytes = 6,5GB >> - flink_taskmanager_Status_JVM_Memory_Heap_Max = 4,7GB >> - flink_taskmanager_Status_JVM_Memory_NonHeap_Used = 325MB >> - flink_taskmanager_Status_JVM_Memory_Direct_MemoryUsed = 500MB >> >> This is my understanding based on all the documentation and observations: >> container_memory_working_set_bytes will be the total amount of memory in >> use, disregarding OS page & block cache. >> Heap will be heap. >> NonHeap is mostly the metaspace. >> Direct_Memory is mostly network buffers. >> >> Running the numbers I have 1 GB unaccounted for. I'm also uncertain as to >> RocksDB. According to the docs RocksDB has a "Column Family Write Buffer" >> where "You need to budget for 2 x your worst case memory use". >> We have 17 ValueStateDescriptors (ignoring state for windows) which I'm >> assuming corresponds to a "Column Family" in RockDB. Meaning our budget >> should be around 2GB. >> Is this accounted for in one of the flink_taskmanager metrics above? >> We've also enabled various rocksdb metrics, but it's unclear where this >> Write Buffer memory would be represented. >> >> Finally, we've seen that when our job has issues and is restarted >> rapidly, NonHeap_Used grows from an initial 50Mb to 700MB, before our >> containers are killed. We're assuming this is due >> to no form of cleanup in the metaspace as classes get (re)loaded. >> >> These are our taskmanager JVM settings: -XX:+UseG1GC >> -XX:MaxDirectMemorySize=1G -XX:+UnlockExperimentalVMOptions >> -XX:+UseCGroupMemoryLimitForHeap -XX:MaxRAMFraction=2 >> With flink config: >> taskmanager.heap.size: 5000m >> state.backend: rocksdb >> state.backend.incremental: true >> state.backend.rocksdb.timer-service.factory: ROCKSDB >> >> Based on what we've observed we're thinking about setting >> -XX:MaxMetaspaceSize to a reasonable value, so that we at least get an >> error message which can easily be traced back to the behavior we're seeing. >> >> Okay, all that said let's sum up what we're asking here: >> - Is there any more insight into how memory is accounted for than our >> current metrics? >> - Which metric, if any accounts for RocksDB memory usage? >> - What's going on with the Metaspace growth we're seeing during job >> restarts, is there something we can do about this such as setting >> -XX:MaxMetaspaceSize? >> - Any other tips to improve reliability running in resource constrained >> environments such as Kubernetes? >> >> Thanks, >> >> William >> >>
