> On April 2, 2014, 3:15 p.m., Josh Elser wrote: > > docs/src/main/resources/design/ACCUMULO-378-design.mdtext, line 62 > > <https://reviews.apache.org/r/19862/diff/1/?file=543190#file543190line62> > > > > Thinking about this from a total ordering standpoint. Say we're > > replicating to two slaves, and we have three rfiles to replicate (1, 2 and > > 3) to those two slaves. > > > > We replicate rfile1 to both, but then the link to slave2 goes down. We > > can still replicate rfile2 and then rfile3 to slave1, while we try to send > > rfile2 to slave2. > > > > What, if instead of the link being down, we happen to communicate to an > > angry server inside of slave2 which never completes the transfer. We don't > > want to transfer rfile3 to attempt to better preserve global ordering. > > > > This can be restated as "we only want to replicate one 'file' to a > > slave at a time" so that we preserve the original semantics of the > > replication "queue" (table). The problem is that this could drastically > > slow down replication when the link between master and slave cannot be > > saturated by one replication task at a time. > > > > This isn't anything that we can reliably guarantee now (without > > conditional mutations), right? Is it worth trying to tackle? The one clear > > change I want to make is that we do want to put the identifier for the > > slave in with the replication record rather than defer determination of > > where a record should be replicated. > > kturner wrote: > I also think transferring files should be an external concern like Mike > Said. One way this could work is the following. > > 1. Cluster A exports a batch of file uris and a control file (similar to > export table) > 2. The user distcps the uris and control file > 3. The control file and dir containing distcp files is provided to > cluter B to import > > The difference between this and import/export table is thats its > stateful. Export on Cluster A provides the list of changes since the last > export. The control file contains ordering information about how to apply > the files. The control file also contains ordering information about other > import/exports. But this process is incomplete. The feedback process would > need to be worked out. The entire process should be resiliant to users > trying to apply things out of order.
What would the ordering guarantees be that we would need to provide for bulk import? Ordering of when bulk-imports finished on the master? If you have two tables and two bulk imports which each had a file to each of those tables, could you have interleaved imports? e.g. table1 sees the file from import1 and then import2, while table2 sees the file from import2 and then import1? I am leaning towards not supporting bulk import replication for simplicity's sake on a first go now. - Josh ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/19862/#review39260 ----------------------------------------------------------- On April 1, 2014, 1:58 a.m., Josh Elser wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/19862/ > ----------------------------------------------------------- > > (Updated April 1, 2014, 1:58 a.m.) > > > Review request for accumulo. > > > Bugs: ACCUMULO-378 > https://issues.apache.org/jira/browse/ACCUMULO-378 > > > Repository: accumulo > > > Description > ------- > > Re-posting a version of the design doc that I own. Contains grammatical fixes > from round one, with a few extra clarifications. New content should be posted > here, but I'll maintain the old review as discussion progresses. > > > Diffs > ----- > > docs/src/main/resources/design/ACCUMULO-378-design.mdtext PRE-CREATION > > Diff: https://reviews.apache.org/r/19862/diff/ > > > Testing > ------- > > > Thanks, > > Josh Elser > >
