Thanks for the write up Stack. I could not make it to Shenzhen. Nice to know the conference and meet up went great.
Regards Ram On Mon, Aug 7, 2017 at 9:36 PM, Stack <[email protected]> wrote: > At fancy Huawei headquarters, 10:00-12:00AM or so (with nice coffee and > fancy little cake squares provided about half way through the session). > > For list of attendees, see picture at end of this email. > > Discussion was mostly in Chinese with about 25% in English plus some > gracious sideline translation so the below is patchy. Hopefully you get the > gist. > > For client-side scanner going against hfiles directly; is there a means of > being able to pass the permissions from hbase to hdfs? > > Issues w/ the hbase 99th percentile were brought up. "DynamoDB can do > 10ms". How to do better? > > SSD is not enough. > > GC messes us up. > > Will the Distributed Log Replay come back to help improve MTTR? We could > redo on new ProcedureV2 basis. ZK timeout is the biggest issue. Do as we > used to and just rely on the regionserver heartbeating... > > Read replica helps w/ MTTR. > > Ratis incubator project to do a quorum based hbase? > > Digression on licensing issues around fb wangle and folly. > > Redo of hbase but quorum based would be another project altogether. > > Decided to go around the table to talk about concerns and what people are > working on. > > Jieshan wondered what could be done to improve OLAP over hbase. > > Client side scanner was brought up again as means of skipping RS overhead > and doing better OLAP. > > Have HBase compact to parquet files. Query parquet and hbase. > > At Huawei, they are using 1.0 hbase. Most problems are assignment. They > have .5M regions. RIT is a killer. Double assignment issues. And RIT. They > run their own services. Suggested they upgrade to get fixes at least. Then > 2.0. > > Will HBase federate like HDFS? Can Master handle load at large scale? It > needs to do federation too? > > Anyone using Bulk loaded replication? (Yes, it just works so no one talks > about it...) > > Request that fixes be backported to all active branches, not just most > current. > > Andrew was good at backporting... not all RMs are. > > Too many branches. What should we do? > > Proliferation of branches makes for too much work. > > Need to cleanup bugs in 1.3. Make it stable release now. > > Lets do more active EOL'ing of branches. 1.1?. > > Hubert asked if we can have clusters where RS are differently capable? > i.e. several generations of HW all running in the same cluster. > > What if fat server goes down. > > Balancer could take of it all. RS Capacity. Balancer can take it into > account. > Regionserver labels like YARN labels. Characteristics. > > Or run it all in docker when heterogeneous cluster. The K8 talk from day > before was mentioned; we should all look at being able to deploy in k8 and > docker. > > Lets put out kubernetes blog...(Doing). > > Alibaba looking at HBase as native YARN app. > > i/o is hard even when containers. > > Use autoscaler of K8 when heavy user. > > Limit i/o use w/ CP. Throttle. > > Spark and client-side scanner came up again. > > Snapshot input format in spark. > > HBase federation came up again. jd.com talking of 3k to 4k nodes in a > cluster. Millions of regions. Region assignment is messing them up. > > Maybe federation is good idea? Argument that it is too much operational > conplexity. Can we fix master load w/ splittable meta, etc? > > Was brought up that even w/ 100s of RS there is scale issue, nvm thousands. > > Alibaba talked about disaster recovery. Described issue where HDFS has > fencing problem during an upgrade. There was no active NN. All RS went down. > ZK is another POF. If ZK is not available. Operators were being asked how > much longer the cluster was going to be down but they could not answer the > question. No indicators from HBase on how much longer it will be down or > how many WALs its processed and how many more to go. Operator unable to > tell his org how long it would be before it all came back on line. Should > say how many regions are online and how many more to do. > > Alibaba use SQL to lower cost. HBase API is low-level. Row-key > construction is tricky. New users make common mistakes. If you don't do > schema right, high-performance is difficult. > > Alibaba are using a subset of Phoenix... simple sql only; throws > exceptions if user tries to do joins, etc.., anything but basic ops. > > HareQL is using hive for meta store. Don't have data typing in hbase. > > HareQL could perhaps contribute some piece... or a module in hbase to > sql... From phoenix? > > Secondary index. > > Client is complicated in phoenix. Was suggested thin client just does > parse... and then offload to server for optimization and execution. > > Then secondary index. Need transaction engine. Consistency of secondary > index. > > We adjourned. > > Your dodgy secretary, > St.Ack > P.S. Please add to this base set of notes if I missed anything. > > > >
