One interesting factor is how load balancer can keep the regions from one region server S1 together in case S1 goes down. I assume all regions from S1 should be moved onto a single region server S2 - otherwise how do we know that related data are still served by the same server ?
I think the above cannot easily be accommodated under the original sense of load balancing. As Todd said in his response, we should carefully consider the implications of new functionality before putting the idea down in code. Sorry for the diversion of the discussion. On Wed, Jan 18, 2012 at 7:03 PM, lars hofhansl <[email protected]> wrote: > Was thinking about that as well. That would be doable. > > Would still need to be some sort of distributed transaction (in the sense > there would be a prepare/vote and commit > phase between the participating regions),but it would all be local to a > single server. > > > > ________________________________ > From: Ted Yu <[email protected]> > To: [email protected]; lars hofhansl <[email protected]> > Sent: Wednesday, January 18, 2012 6:51 PM > Subject: Re: Limited cross row transactions > > Still need to go over the patch, Lars. > > I wonder how difficult supporting cross-region transactions in the same > region server would be. > > Cheers > > On Wed, Jan 18, 2012 at 5:02 PM, lars hofhansl <[email protected]> > wrote: > > > Filed https://issues.apache.org/jira/browse/HBASE-5229 for further > > discussion, attached a patch that does this. > > > > > > As for your point... > > The below is one way to define limited groups of rows that can > participate > > in transactions (I should not have named it parent/child, that just > > confuses my point). > > Your scenario calls for global transaction (unless you have to some other > > approach to limit the scope of rows that could participate in your FK > > transactions to something less than the entire database). > > > > If every transaction is a global transaction the database will not scale. > > > > See http://www.julianbrowne.com/article/viewer/brewers-cap-theorem > > and > > > http://www.cloudera.com/blog/2010/04/cap-confusion-problems-with-partition-tolerance/ > > > > Also check out two phase commit failure and blocking scenarios, and > Paxos' > > conditions for termination. > > > > -- Lars > > > > > > ----- Original Message ----- > > From: Mikael Sitruk <[email protected]> > > To: [email protected]; lars hofhansl <[email protected]> > > Cc: > > Sent: Wednesday, January 18, 2012 12:01 AM > > Subject: Re: Limited cross row transactions > > > > This is for parent child relationship, but what if there is no parent > child > > relationship, but more a foreign key like relationship? > > Using this model you do a full scan to get all the index (since you don't > > know the parent, you just know the "secondary index"). > > Or will you use a group ID as a prefix of parent key and "child" key? In > > this case splitting according to group may be more difficult, (due to > > different growth of groups). > > Doing this aren't we back in the headache of sharding in rdbms? > > > > Mikael.S > > > > > > On Wed, Jan 18, 2012 at 7:45 AM, lars hofhansl <[email protected]> > > wrote: > > > > > This thread is probably getting too long... > > > > > > In HBase we have to let go of "global stuff". I submit that global > > > transactions across 1000's of nodes that can fail will never work > > > adequately. > > > For that kind of consistency you will be hit in availability. > > > > > > Like Megastore the trick is in creating a local grouping of entities > that > > > can participate in local transactions. > > > If you limit the (consistent) index to child entities of parent entity > > you > > > can form your index like this: > > > parentKey1... > > > parentKey1.childTableName1.indexedField1 > > > parentKey1.childTableName1.indexedField2 > > > ... > > > parentKey1.childTableName2.indexedField1 > > > parentKey1.childTableName2.indexedField2 > > > ... > > > (assuming . cannot be in any parent key or child table name here, but > you > > > get the idea). > > > > > > > > > When scanning the parent you'd have to skip the index rows with a > filter. > > > Within a parentKey you can find childKeys efficiently by scanning the > > > index rows. > > > > > > Since the parent and the index entries would sort together the table > can > > > be pre-split (or one could have a simple prefix based balancer). > > > > > > -- Lars > > > > > > ----- Original Message ----- > > > From: Mikael Sitruk <[email protected]> > > > To: [email protected] > > > Cc: > > > Sent: Tuesday, January 17, 2012 3:07 PM > > > Subject: Re: Limited cross row transactions > > > > > > Well i understand the limitation now, asking to be in the same region > is > > > really hard constraint. > > > Even if this is on the same RS this is not enough, because after a > > restart, > > > regions may be allocated differently and now part of the data may be in > > one > > > region under server A and the other part under server B. > > > > > > Well perhaps we need use case for better understanding, and perhaps > > finding > > > alternative. > > > > > > The first use case i was thinking of is as follow - > > > I need to insert data with different access criteria, but the data > > inserted > > > should be inserted in atomic way. > > > In RDBMS i would have two table, insert data in the first one with > key#1 > > > and then in the second one with key #2 then commit. > > > In HBase i need to use different column family with key #1 (for > > atomicity) > > > then to manage a kind of secondary index to map key#2 to key #1 > (perhaps > > > via co-processor) to have quick access to the data of key#2. > > > Having cross row trx, i would think of sing different keys under the > same > > > table (and probably different cf too), without the need to have > secondary > > > index, but again with the limitation it does not seems to be easily > > > feasible. > > > > > > Mik. > > > > > > On Wed, Jan 18, 2012 at 12:22 AM, Ted Yu <[email protected]> wrote: > > > > > > > People rely on RDBMS for the transaction support. > > > > > > > > Consider the following example: > > > > A highly de-normalized schema puts related users in the same region > > where > > > > this 'limited cross row transactions' works. > > > > After some time, the region has to be split (maybe due to good > business > > > > condition). > > > > What should the HBase user do now ? > > > > > > > > Cheers > > > > > > > > On Tue, Jan 17, 2012 at 2:13 PM, Mikael Sitruk < > > [email protected] > > > > >wrote: > > > > > > > > > Ted - My 2 cents as a user. > > > > > The user should know what he is doing, this is like a 'delete' > > > operation, > > > > > this is less intuitive that the original delete in RDBMS, so the > same > > > > will > > > > > be for this light transaction. > > > > > If the transaction fails because of cross region server then the > > design > > > > of > > > > > the user was wrong > > > > > if the transaction fails because of concurrent access, then he > should > > > be > > > > > able to re-read and reprocess its request. > > > > > The only problem is how to make sure in advance that the different > > rows > > > > > will be in the same RS? > > > > > > > > > > Lars - is the limitation is at the region or at the region server? > It > > > was > > > > > not so clear. > > > > > > > > > > Mikael.S > > > > > > > > > > On Tue, Jan 17, 2012 at 11:52 PM, Ted Yu <[email protected]> > > wrote: > > > > > > > > > > > Back to original proposal: > > > > > > If client side grouping reveals that the batch of operations > cannot > > > be > > > > > > supported by 'limited cross row transactions', what should the > user > > > do > > > > ? > > > > > > > > > > > > Cheers > > > > > > > > > > > > On Tue, Jan 17, 2012 at 1:49 PM, Ted Yu <[email protected]> > > wrote: > > > > > > > > > > > > > Whether Omid fits the bill is open to discussion. > > > > > > > > > > > > > > We should revisit HBASE-2315 and provide the support Flavio, et > > al > > > > > need. > > > > > > > > > > > > > > Cheers > > > > > > > > > > > > > > > > > > > > > On Tue, Jan 17, 2012 at 1:41 PM, Lars George < > > > [email protected] > > > > > > >wrote: > > > > > > > > > > > > > >> Hi Ted, > > > > > > >> > > > > > > >> Wouldn't Omid (https://github.com/yahoo/omid) help there? Or > is > > > > that > > > > > > too > > > > > > >> broad? Just curious. > > > > > > >> > > > > > > >> Lars > > > > > > >> > > > > > > >> On Jan 17, 2012, at 4:36 PM, Ted Yu wrote: > > > > > > >> > > > > > > >> > Can we collect use case for 'limited cross row transactions' > > > > first ? > > > > > > >> > > > > > > > >> > I have been thinking about (unlimited) multi-row transaction > > > > support > > > > > > in > > > > > > >> > HBase. It may not be a one-man task. But we should > definitely > > > > > > implement > > > > > > >> it > > > > > > >> > someday. > > > > > > >> > > > > > > > >> > Cheers > > > > > > >> > > > > > > > >> > On Tue, Jan 17, 2012 at 1:27 PM, lars hofhansl < > > > > [email protected] > > > > > > > > > > > > >> wrote: > > > > > > >> > > > > > > > >> >> I just committed HBASE-5203 (together with HBASE-3584 this > > > > > implements > > > > > > >> >> atomic row operations). > > > > > > >> >> Although a relatively small patch it lays the groundwork > for > > > > > > >> heterogeneous > > > > > > >> >> operations in a single WALEdit. > > > > > > >> >> > > > > > > >> >> The interesting part is that even though the code enforced > > the > > > > > atomic > > > > > > >> >> operation to be a for single row, this is not required. > > > > > > >> >> It is enough if all involved KVs reside in the same region. > > > > > > >> >> > > > > > > >> >> I am not saying that we should add any high level concept > to > > > > HBase > > > > > > >> (such > > > > > > >> >> as the EntityGroups of Megastore). > > > > > > >> >> > > > > > > >> >> But, with a slight addition to the API (allowing a grouping > > of > > > > > > multiple > > > > > > >> >> row operations) client applications have all the building > > > blocks > > > > to > > > > > > do > > > > > > >> >> limited cross row atomic operations. > > > > > > >> >> The client application would be responsible for either > > > correctly > > > > > > >> >> pre-splitting the table, or a custom balancer has to be > > > provided. > > > > > > >> >> > > > > > > >> >> The operation would fail if the regionserver determines > that > > it > > > > > would > > > > > > >> need > > > > > > >> >> data from multiple region servers. > > > > > > >> >> > > > > > > >> >> I think this needs at least minimal support from HBase and > > > cannot > > > > > > >> >> (efficiently or without adding more moving parts) by a > client > > > API > > > > > > only. > > > > > > >> >> > > > > > > >> >> > > > > > > >> >> Comments? Is this worth pursuing? If so, I'll file a jira > and > > > > > > provide a > > > > > > >> >> patch. > > > > > > >> >> > > > > > > >> >> Thanks. > > > > > > >> >> > > > > > > >> >> > > > > > > >> >> -- Lars > > > > > > >> >> > > > > > > >> >> > > > > > > >> > > > > > > >> > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > Mikael.S > > > > > > > > > > > > > > > > > > > > > -- > > > Mikael.S > > > > > > > > > > > > -- > > Mikael.S > > > > >
