Hi Andrew, As we described more in http://yahooeng.tumblr.com/post/116291838351/pistachio-co-locate-the-data-and-compute-for, a very common problem we saw in Hadoop use cases is we often need to persist the previous result of one map reduce job onto HDFS, then the next day we process the new data together with the previous result. Usually the most expensive part is the shuffling part where we need to join the previous data and the new data together. It's so expensive because HDFS doesn't store the data in a partitioned way. So data have to be transferred again and again in the shuffling phase. Instead, in Pistachio we do the computation right on top of the partitioned storage layer, so that the previous result is always stored in a partitioned way, so shuffling can be avoided. Expensive IO and roundtrips can thus be avoided so that much better performance can be achieved.
The other difference is in Pistachio we can do computation based on in-memory storage with data replication. Different from the in-memory computation in Spark, the storage can be in-memory here. Please let me know if I'm not clear enough. Thanks, Gavin Li On Mon, Jun 22, 2015 at 7:53 PM, Andrew Purtell <apurt...@apache.org> wrote: > It was a simple question, and not meant to suggest anything one way or > other regarding my opinion of this proposal. > > On Monday, June 22, 2015, John D. Ament <johndam...@apache.org> wrote: > > > On Mon, Jun 22, 2015 at 10:26 PM Andrew Purtell <apurt...@apache.org > > <javascript:;>> wrote: > > > > > > Pistachio can easily embed computation to the storage layer to > achieve > > > the > > > > best data locality to improve the computation performance > significantly > > > > which is an innovative model comparing with the normal ways where the > > > > storage and compute are independent to each other. > > > > > > Have you heard of something called Hadoop? > > > > > > > Regardless of whether he has or not - what's your point? The ASF has > > historically not denied the entry of new projects just because their > domain > > intersects with another project's. > > > > > > > > > > > > > On Thu, Jun 18, 2015 at 10:17 AM, Gavin Li <lyo.ga...@gmail.com > > <javascript:;>> wrote: > > > > > > > Hi, > > > > > > > > I want to propose project Pistachio to enter Apache Incubator. > > > > > > > > Below please find the proposal. > > > > > > > > Thanks, > > > > Gavin Li > > > > > > > > > > > > > > > > = Pistachio = > > > > > > > > == Abstract == > > > > > > > > Pistachio is a fault-tolerant low latency distributed storage system > > > which > > > > enables simple embedding the computation to the storage layer to > > achieve > > > > best data locality. It evolves from Yahoo’s global user profile > storage > > > > system. > > > > > > > > == Proposal == > > > > > > > > Pistachio is a distributed key value store system with fault > tolerance > > > and > > > > consistency guarantee. It supports multiple local storage engine > > > including > > > > in-memory, kyoto cabinet, rocks DB etc. Pistachio is being used as > the > > > user > > > > profile storage for massive scale global ads products in Yahoo > storing > > > 10+ > > > > billion user profiles. The performance and reliability has been well > > > proven > > > > on production. > > > > > > > > Pistachio can easily embed computation to the storage layer to > achieve > > > the > > > > best data locality to improve the computation performance > significantly > > > > which is an innovative model comparing with the normal ways where the > > > > storage and compute are independent to each other. > > > > > > > > == Background == > > > > > > > > Pistachio is originally designed and optimized for Yahoo’s large > scale > > > > global open RTB(real-time bidding) use cases where latency is > > > critical(the > > > > whole request needs to be finished within 100ms including network > round > > > > trips). It stores 10+ billion user profiles in 8 data centers. > > > > > > > > Then because of the great performance and the flexibility of local > > > storage > > > > choices, we evolved it to do distributed compute. Rich call back > > > interfaces > > > > are added to supports easy compute directly on top of the storage > > system > > > > local to the data partition. This model is totally different from the > > > > traditional distributed computation model where the storage and > compute > > > are > > > > separated and independent. In the new model we found data locality > can > > be > > > > improved significantly and lots of data access round trips can be > > reduced > > > > in computation, and the performance can be improved significantly. > > > > > > > > It was publicly announced in April 2015 and currently being hosted in > > > > Github. > > > > > > > > == Rationale == > > > > > > > > As a key value store system Pistachio is unique in terms of low > latency > > > > access with fault tolerance and consistency guarantee. The > reliability, > > > > scalability, fault tolerance and performance has been well proven in > > > global > > > > large scale revenue supporting production system in Yahoo. > > > > > > > > As a distributed computation system, it’s an innovative model where > the > > > > compute layer is introduced on top of the storage layer natively and > > > > naturally to optimize the data locality of computation. > > > > > > > > Operating the project in “apache way” greatly aligns with the > long-term > > > > vision of this project and can greatly help the development of the > > > > community. > > > > > > > > == Current Status == > > > > > > > > Pistachio was open-sourced and announced in April 2015 and currently > > > being > > > > hosted in Github, it was mainly being developed by the team from > Yahoo > > > and > > > > already attracted lots of external developers (20+ watches and forks > on > > > > github). > > > > > > > > == Meritocracy == > > > > > > > > We plan to build an environment following the Apache meritocracy > > > > principles. Many companies including Linkedin, GF securities, > Microsoft > > > and > > > > open source communities like deeplearning4j have already expressed > > > > interests or accepted the invitations to participate in this project. > > > > > > > > == Community == > > > > > > > > Since the announcement of Pistachio we received lots of interests. > And > > > the > > > > concept of embedding computation to storage also got lots of > > > recognitions. > > > > We also started to work with other communities like deeplearning4j to > > > build > > > > more application use cases with Pistachio. We believe the community > > will > > > > grow fast. > > > > > > > > == Core Developers == > > > > > > > > This project is created by Gavin Li. Core developers are currently > > mainly > > > > in Yahoo. > > > > > > > > == Alignment == > > > > > > > > Pistachio depends on many Apache projects and dependencies including > > > Kafka, > > > > Helix, Zookeeper, Curator, Apache Commons, etc. > > > > > > > > == Known Risks == > > > > > > > > === Orphaned Products === > > > > > > > > The risk of Pistachio being orphaned is small because Yahoo heavily > > > > invested in this system. It’s the internal storage standard for > Yahoo’s > > > > global ads products and still being expanded. Migration cost from > this > > > > project is very high. We are also working with external communities > > like > > > > deeplearning4j and other companies to expand the applications. > > > > > > > > === Inexperience with Open Source === > > > > > > > > Core developers are experienced open source contributors in many > > projects > > > > including Druid, Spark, Storm, etc. Pistachio committers will be > guided > > > by > > > > the mentors with strong Apache open source project backgrounds. > > > > > > > > === Homogeneous Developers === > > > > > > > > The initial committers include developers from several institutions > > > > including Microsoft, GF Securities, Linkedin and Yahoo. > > > > > > > > === Reliance on Salaried Developers === > > > > > > > > We work on Pistachio on both salaried time and after hours. Many > > > developers > > > > from other institutions already accepted the invitation to volunteer > > > > working on Pistachio. > > > > > > > > === Relationships with Other Apache Products === > > > > > > > > As mentioned earlier, Pistachio depends on apache kafka, helix, > > > zookeeper, > > > > curator, etc. > > > > > > > > === A Excessive Fascination with the Apache Brand === > > > > > > > > Generating publicity is not the purpose of this proposal. We mainly > > want > > > to > > > > join the ASF in order to increase our contacts and visibility in the > > open > > > > source world to attract great developers. > > > > > > > > == Document == > > > > > > > > Current documentation can be found here: > > > > https://github.com/yahoo/Pistachio. > > > > > > > > == Initial source == > > > > > > > > Initial source can be found here in the Github repo: > > > > https://github.com/yahoo/Pistachio. > > > > > > > > == External dependencies == > > > > > > > > To the best of our knowledge, here is the list of dependencies: > > > > Rocks DB > > > > ICU4j > > > > Apache Curator > > > > netty > > > > google http client > > > > codahale.metrics > > > > apache helix > > > > apache zookeeper > > > > apache commons > > > > apache thrift > > > > apache kafka > > > > kyoto cabinet (GNU GPL) > > > > google protocol buffer > > > > kryo > > > > slf4j > > > > > > > > To the best of our knowledge, except kyoto cabinet others are all > > > > distributed under Apache compatible licenses: > > > > BSD > > > > ICU > > > > Apache License 2.0 > > > > MIT > > > > > > > > Kytoto cabinet is under GNU GPL, but it is not a hard necessary > > > dependency > > > > to Pistachio, it’s an optional pluggable storage engine. It’s > designed > > in > > > > the way that it’s totally plugable and very loosely coupled. We can > > > easily > > > > remove it in graduation. > > > > > > > > == Required Resources == > > > > > > > > Mailing Lists > > > > > > > > pistachio-user > > > > pistachio-dev > > > > pistachio-commits > > > > pistachio-private (for private PMC discussions) > > > > > > > > Git > > > > > > > > The Pistachio team prefers Git for source version control: git:// > > > > git.apache.org/pistachio > > > > > > > > Issue Tracking > > > > > > > > JIRA Pistachio (PISTACHIO) > > > > > > > > Other Resources > > > > > > > > Jenkins continuous integration testing > > > > > > > > == Initial Committers == > > > > > > > > Gavin Li <lyo.gavin at gmail dot com> > > > > Lie Yang <lyang at yahoo-inc dot com> > > > > Jay Kim <pitecus at yahoo-inc dot com> > > > > Flavio Junqueira <fpj at apache dot org> > > > > Chihong Liang<chihong.liang at gmail dot com> > > > > Yong Liu<ly7110 at gmail dot com> > > > > Shengwu Yang <yangshengwu at gmail dot com> > > > > > > > > == Affiliations == > > > > > > > > Gavin Li - Yahoo > > > > Flavio Junqueira - Microsoft > > > > Chihong Liang - GF securities > > > > Yong Liu - Yingmi Asset Management Corp. > > > > Lie Yang - Yahoo > > > > Jay Kim - Yahoo > > > > Shengwu Yang - Linkedin China > > > > > > > > == Sponsors == > > > > > > > > === Champion === > > > > > > > > Flavio Junqueira <fpj at apache dot org> > > > > > > > > === Nominated Mentors === > > > > > > > > === Sponsoring Entity === > > > > > > > > The Apache Incubator > > > > > > > > > > > > > > > > -- > > > Best regards, > > > > > > - Andy > > > > > > Problems worthy of attack prove their worth by hitting back. - Piet > Hein > > > (via Tom White) > > > > > > > > -- > Best regards, > > - Andy > > Problems worthy of attack prove their worth by hitting back. - Piet Hein > (via Tom White) >
