So if the HDFS has a replication factor of m and an r-file has a range that intersects n tablets, then data-locality will never be achieved for max(0,n-m) of the r-files, that is, they will never be on the same node as their tablet server until compaction, correct?
-- Jeff Kubina 410-988-4436 On Wed, Oct 7, 2015 at 9:35 AM, Josh Elser <josh.el...@gmail.com> wrote: > > On Oct 7, 2015 8:47 AM, "Jeff Kubina" <jeff.kub...@gmail.com> wrote: > > > > How does Accumulo process an r-file for bulk ingesting when the key > range of an r-file is within one tablet's key range and when the key range > of an r-file spans two or more tablets? > > > > If the r-file is within one tablet's range I thought the file was "just > renamed" and added to the tablet's list of r-files. Is that correct? > > Bingo > > > If the key range of the r-file spans two or more files is the r-file > partitioned into separate r-files for each appropriate tablet server or are > the records "batch-written" to each appropriate tablet in memory? > > They're logically partitioned if memory serves (the files are not > rewritten). So you would see multiple entries in the metadata table for a > single file with certain offsets. No replaying of mutations by batch > writers. >
