Billie, I think you've got it. Now I need to write it.
On Thu, Jul 12, 2012 at 11:47 AM, Billie J Rinaldi <[email protected]> wrote: > On Thursday, July 12, 2012 8:47:41 AM, "David Medinets" > <[email protected]> wrote: >> I'd like to track field level changes for a given record (say, >> author). So I create a table without a VersioningIterator. And I >> insert a few records: >> >> insert "JOHN" "ATTRIBUTE" "AGE" "34" >> insert "JOHN" "ATTRIBUTE" "HEIGHT" "67" >> insert "JOHN" "BOOKS" "TITLE" "THE RISE OF ACCUMULO" >> >> The next action is that some ingest process happens and does this: >> >> insert "JOHN" "ATTRIBUTE" "AGE" "34" >> >> Since there is no VersioningIterator, there are two AGES both with >> "34" as the value. >> >> I would like an DropUnchangedValueIterator which removes the last >> inserted record. Removing the last record lets me use the n-1 >> timestamp as a LastUpdated value for the key-value pair. But as soon >> as a record is deleted, the previous records are not available >> anymore? What if the timestamp is set to MAX-timestamp so the records >> are sorted backwards? Does that avoid the blocking tombstones? I'd >> look at the source code before asking but I don't have that luxury for >> the next week or two and the question is rattling around my head. > > This is mixing the idea of a deletion entry, which removes all earlier > entries, and the the idea that iterators can arbitrarily filter out entries. > I don't think reversing the timestamp will help you much in this case; what > you want is an iterator that does pairwise comparisons of entries, and if the > values are the same keep one entry with the earlier timestamp (then keep > comparing entries for that record), and if the values are different keep one > entry with the later timestamp (then skip to the next record). I think > you'll have to write a custom iterator for that. > > Billie > > >> Naturally, I could query the database before the ingest insert. But, >> referring to slide 19 in Adam's presentation at >> http://people.apache.org/~afuchs/slides/accumulo_table_design.pdf, the >> read-modify-write design is not optimal.
