Hi Ravi, I took a quick look at Hazlecast, what they offer is a distributed map across cluster (on any single node only portion of the map is stored), to facilitate parallel data loading I think we need a complete copy on each node, is this the structure we are looking for?
it does allow map in-memory backup in case one node goes down, to ensure its persistency, they allow storing map to db, but requires implementing their API to hook them up, there are async/ sync mode supported with no guarantee in terms of consistency, unless going further for a transaction support, 2-phase commit/XA are offered with read-committed isolation, to achieve that is quite complicated when we need to ensure ACID on changes to the map. I suggest you to investigate further to understand the implication and effort. We all understand We couldn't afford any inconsistency on dictionary, that means we couldn't decode the data back correctly. correctness is even more critical compared to performance. Jihong -----Original Message----- From: Ravindra Pesala [mailto:ravi.pes...@gmail.com] Sent: Saturday, October 15, 2016 12:50 AM To: dev Subject: Re: Discussion(New feature) regarding single pass data loading solution. Hi Jacky/Jihong, I agree that new dictionary values are less in case of incremental data load but that is completely depends on user data scenarios. In some user scenarios new dictionary values may be more we cannot overrule that. And also for users convenience we should provide single pass solution with out insisting them to run external tool first. We can provide the option to run external tool first and provide dictionary to improve performance. My opinion is better to use some professional distributed map like Hazlecast than Zookeeper + HDFS. It is lite weight and does not require to have separate cluster, it can form the cluster within the executor jvm's . May be we can have a try, after all it will be just one interface implementation for dictionary generation. We can have multiple implementations and then decide based on optimal performance. Regards, Ravi On 15 October 2016 at 10:50, Jacky Li <jacky.li...@qq.com> wrote: > Hi, > > I can offer one more approach for this discussion, since new dictionary > values are rare in case of incremental load (ensure first load having as > much dictionary value as possible), so synchronization should be rare. So > how about using Zookeeper + HDFS file to provide this service. This is what > carbon is doing today, we can wrap Zookeeper + HDFS to provide the global > dictionary interface. > It has the benefit of > 1. automated: without bordering the user > 2. not introducing more dependency: we already using zookeeper and HDFS. > 3. performance? since new dictionary value and synchronization is rare. > > What do you think? > > Regards, > Jacky > > > 在 2016年10月15日，上午2:38，Jihong Ma <jihong...@huawei.com> 写道： > > > > Hi Ravi, > > > > The major concern I have for generating global dictionary from scratch > with a single scan is performance, the way to handle an occasional update > to the dictionary is way simpler and cost effective in terms of > synchronization cost and refresh the global/local cache copy. > > > > There are a lot to worry about for distributed map, and leveraging KV > store is overkill if simply just for dictionary generation. > > > > Regards. > > > > Jihong > > > > -----Original Message----- > > From: Ravindra Pesala [mailto:ravi.pes...@gmail.com] > > Sent: Friday, October 14, 2016 11:03 AM > > To: dev > > Subject: Re: Discussion(New feature) regarding single pass data loading > solution. > > > > Hi Jihong, > > > > I agree, we can use external tool for first load, but for incremental > load > > we should have solution to add global dictionary. So this solution should > > be enough to generate global dictionary even if user does not use > external > > tool for first time. That solution could be distributed map or KV store. > > > > Regards, > > Ravi. > > > > On 14 October 2016 at 23:12, Jihong Ma <jihong...@huawei.com> wrote: > > > >> Hi Liang, > >> > >> This tool is more or less like the first load, the first time after > table > >> is created, any subsequent loads/incremental loads will proceed and is > >> capable of updating the global dictionary when it encounters new value, > >> this is easiest way of achieving 1 pass data loading process without too > >> much overhead. > >> > >> Since this tool is only triggered once per table, not considered too > much > >> burden on the end users. Making global dictionary generation out of the > way > >> of regular data loading is the key here. > >> > >> Jihong > >> > >> -----Original Message----- > >> From: Liang Chen [mailto:chenliang6...@gmail.com] > >> Sent: Thursday, October 13, 2016 5:39 PM > >> To: dev@carbondata.incubator.apache.org > >> Subject: RE: Discussion(New feature) regarding single pass data loading > >> solution. > >> > >> Hi jihong > >> > >> I am not sure that users can accept to use extra tool to do this work, > >> because provide tool or do scan at first time per table for most of > global > >> dict are same cost from users perspective, and maintain the dict file > also > >> be same cost, they always expecting that system can automatically and > >> internally generate dict file during loading data. > >> > >> Can we consider this: > >> first load: make scan to generate most of global dict file, then copy > this > >> file to each load node for subsequent loading > >> > >> Regards > >> Liang > >> > >> > >> Jihong Ma wrote > >>>>>>> the question is what would be the default implementation? Load data > >> without dictionary? > >>> > >>> My thought is we can provide a tool to generate global dictionary using > >>> sample data set, so the initial global dictionaries is available before > >>> normal data loading. We shall be able to perform encoding based on > that, > >>> we only need to handle occasionally adding entries while loading. For > >>> columns specified with global dictionary encoding, but dictionary is > not > >>> placed before data loading, we error out and direct user to use the > tool > >>> first. > >>> > >>> Make sense? > >>> > >>> Jihong > >>> > >>> -----Original Message----- > >>> From: Ravindra Pesala [mailto: > >> > >>> ravi.pesala@ > >> > >>> ] > >>> Sent: Thursday, October 13, 2016 1:12 AM > >>> To: dev > >>> Subject: Re: Discussion(New feature) regarding single pass data loading > >>> solution. > >>> > >>> Hi Jihong/Aniket, > >>> > >>> In the current implementation of carbondata we are already handling > >>> external dictionary while loading the data. > >>> But here the question is what would be the default implementation? Load > >>> data with out dictionary? > >>> > >>> > >>> Regards, > >>> Ravi > >>> > >>> On 13 October 2016 at 03:50, Aniket Adnaik < > >> > >>> aniket.adnaik@ > >> > >>> > wrote: > >>> > >>>> Hi Ravi, > >>>> > >>>> 1. I agree with Jihong that creation of global dictionary should be > >>>> optional, so that it can be disabled to improve the load performance. > >>>> User > >>>> should be made aware that using global dictionary may boost the query > >>>> performance. > >>>> 2. We should have a generic interface to manage global dictionary when > >>>> its > >>>> from external sources. In general, it is not a good idea to depend on > >> too > >>>> many external tools. > >>>> 3. May be we should allow user to generate global dictionary > separately > >>>> through SQL command or similar. Something like materialized view. This > >>>> means carbon should avoid using local dictionary and do late > >>>> materialization when global dictionary is present. > >>>> 4. May be we should think of some ways to create global dictionary > >> lazily > >>>> as we serve SELECT queries. Implementation may not be that straight > >>>> forward. Not sure if its worth the effort. > >>>> > >>>> Best Regards, > >>>> Aniket > >>>> > >>>> > >>>> On Tue, Oct 11, 2016 at 7:59 PM, Jihong Ma < > >> > >>> Jihong.Ma@ > >> > >>> > wrote: > >>>> > >>>>> > >>>>> A rather straight option is allow user to supply global dictionary > >>>>> generated somewhere else or we build a separate tool just for > >>>> generating > >>>> as > >>>>> well updating dictionary. Then the general normal data loading > process > >>>> will > >>>>> encode columns with local dictionary if not supplied. This should > >>>> cover > >>>>> majority of cases for low-medium cardinality column. For the cases we > >>>> have > >>>>> to incorporate online dictionary update, use a lock mechanism to sync > >>>> up > >>>>> should serve the purpose. > >>>>> > >>>>> In another words, generating global dictionary is an optional step, > >>>> only > >>>>> triggered when needed, not a default step as we do currently. > >>>>> > >>>>> Jihong > >>>>> > >>>>> -----Original Message----- > >>>>> From: Ravindra Pesala [mailto: > >> > >>> ravi.pesala@ > >> > >>> ] > >>>>> Sent: Tuesday, October 11, 2016 2:33 AM > >>>>> To: dev > >>>>> Subject: Discussion(New feature) regarding single pass data loading > >>>>> solution. > >>>>> > >>>>> Hi All, > >>>>> > >>>>> This discussion is regarding single pass data load solution. > >>>>> > >>>>> Currently data is loading to carbon in 2 pass/jobs > >>>>> 1. Generating global dictionary using spark job. > >>>>> 2. Encode the data with dictionary values and create carbondata > >> files. > >>>>> This 2 pass solution has many disadvantages like it needs to read the > >>>> data > >>>>> twice in case of csv files input or it needs to execute dataframe > >> twice > >>>> if > >>>>> data is loaded from dataframe. > >>>>> > >>>>> In order to overcome from above issues of 2 pass dataloading, we can > >>>> have > >>>>> single pass dataloading and following are the alternate solutions. > >>>>> > >>>>> Use local dictionary > >>>>> Use local dictionary for each carbondata file while loading data, but > >>>> it > >>>>> may lead to query performance degradation and more memory footprint. > >>>>> > >>>>> Use KV store/distributed map. > >>>>> *HBase/Cassandra cluster : * > >>>>> Dictionary data would be stored in KV store and generates the > >>>> dictionary > >>>>> value if it is not present in it. We all know the pros/cons of Hbase > >>>> but > >>>>> following are few. > >>>>> Pros : These are apache licensed > >>>>> Easy to implement to store/retreive dictionary values. > >>>>> Performance need to be evaluated. > >>>>> > >>>>> Cons : Need to maintain seperate cluster for maintaining global > >>>>> dictionary. > >>>>> > >>>>> *Hazlecast distributed map : * > >>>>> Dictionary data could be saved in distributed concurrent hash map of > >>>>> hazlecast. It is in-memory map and partioned as per number of nodes. > >>>> And > >>>>> even we can maintain the backups using sync/async functionality to > >>>> avoid > >>>>> the data loss when instance is down. We no need to maintain seperate > >>>>> cluster for it as it can run on executor jvm itself. > >>>>> Pros: It is apache licensed. > >>>>> No need to maintain seperate cluster as instances can run in > >>>>> executor jvms. > >>>>> Easy to implement and store/retreive dictionary values. > >>>>> It is pure java implementation. > >>>>> There is no master/slave concept and no single point failure. > >>>>> > >>>>> Cons: Performance need to be evaluated. > >>>>> > >>>>> *Redis distributed map : * > >>>>> It is also in-memory map but it is coded in c language so we > >> should > >>>>> have java client libraries to interact with redis. Need to maintain > >>>>> seperate cluster for it. It also can partition the data. > >>>>> Pros : More feature rich than Hazlecast. > >>>>> Easy to implement and store/retreive dictionary values. > >>>>> Cons : Need to maintain seperate cluster for maintaining global > >>>>> dictionary. > >>>>> May not be suitable for big data stack. > >>>>> It is BSD licensed (Not sure whether we can use or not) > >>>>> Online performance figures says it is little slower than hazlecast. > >>>>> > >>>>> Please let me know which would be best fit for our loading solution. > >>>> And > >>>>> please add any other suitable solution if I missed. > >>>>> -- > >>>>> Thanks & Regards, > >>>>> Ravi > >>>>> > >>>> > >>> > >>> > >>> > >>> -- > >>> Thanks & Regards, > >>> Ravi > >> > >> > >> > >> > >> > >> -- > >> View this message in context: http://apache-carbondata- > >> mailing-list-archive.1130556.n5.nabble.com/Discussion-New- > >> feature-regarding-single-pass-data-loading-solution-tp1761p1887.html > >> Sent from the Apache CarbonData Mailing List archive mailing list > archive > >> at Nabble.com. > >> > > > > > > > > -- > > Thanks & Regards, > > Ravi > > > > -- Thanks & Regards, Ravi
