The way I picture it, we only need to check the 100 or 1000 regions to get the very first row of a order-by scan. Afterwards, we just need to get the next row from the region that we took the last row from, compare with the existing rows we have, and return the lowest.
So if we have R regions to grab from, the first call to next() will have to fetch all R rows from the regions and do R log R comparisons to do the sort. Then each call to next() will cost 1 row fetch from region plus log R comparisons to put in a sorted tree. On Tue, Apr 22, 2008 at 1:10 PM, Bryan Duxbury <[EMAIL PROTECTED]> wrote: > If you just want the first 10 by a certain prefix ordered by a column, then > you will definitely be better off scanning them by row and ordering them > clientside. > > Surely your idea of maintaining a separate SortedMap in each region would > work, but I don't think you should discount the cost associated with having > to talk to a bunch of different regions every time you want to know what the > next row is. You'll do a lot of extra work to get that merged view of the > index, and potentially the approach won't scale up to queries that might > cover more than a "few" regions - can you imagine having to check 100 or > 1000 regions for the next result every time you needed to iterate? > > > > On Apr 22, 2008, at 12:58 PM, Clint Morgan wrote: > > > > Yeah, that would be an easy approach. We would need HBASE-82. > > > > The main problem I see here is that we cannot take (as much) advantage > > of our row key prefix in weeding out rows. > > > > Say I want the first 10 rows that start with XXX ordered by A:amount, > > then I would have scan through column values from rows everywhere in > > the table until I hit 10 with my prefix. Could be costly if table is > > large compared to the number of rows that start with XXX. > > > > Whereas if we have one SortedMap per Region, then I can quickly narrow > > down to (hopefully) a few regions based on key prefix. > > > > Though other usage / table loading patterns would surely benefit from > > this approach... > > > > On Tue, Apr 22, 2008 at 12:31 PM, Bryan Duxbury <[EMAIL PROTECTED]> wrote: > > > > > This doesn't have to be all that complicated. > > > > > > Why not keep another HBase table as the index? The keys would be the > column > > > values. There'd be a single matchingRows column family, and the > qualifiers > > > and values would be the rows that match that column. Then, when you want > to > > > scan in column order instead of row order, you scan the index table, > find > > > the list of rows that match each column, and then do a random read to > grab > > > those individually. It'll for sure be slower than scanning a table > ordered > > > by rows, but it'll get you what you want. It'll also handle the case > where > > > the column values aren't unique. > > > > > > If you need custom sorting for those values, then HBASE-82 would solve > that > > > problem. > > > > > > -Bryan > > > > > > > > > > > > On Apr 22, 2008, at 11:58 AM, stack wrote: > > > > > > > > > > > > > Some questions interlaced below: > > > > > > > > Clint Morgan wrote: > > > > > > > > > > > > > All, > > > > > > > > > > We want to put secondary indexes into hbase. The motivation is that > we > > > > > are storing data in hbase that we want to serve to users. We would > > > > > like to be able to serve rows sorted by column values. Our queries > > > > > will be over rows with a given key prefix, so we should not be > hitting > > > > > to many regions. > > > > > I was thinking it would work roughly like this: > > > > > > > > > > - At table creation time, individual columns can be declared as > > > > > indexed. By default we could sort the column values > lexicographically, > > > > > or we can provide a WritableComparatorFactory<T> which has the > ability > > > > > to make values of type T from a byte [], as well as providing a > > > > > Comparator<T>. (Better than providing a Comparator<byte[]> as it > only > > > > > costs once per row insert for deserialization, rather that twice on > > > > > each comparison). > > > > > > > > > > > > > > > > > > > > > > > I don't follow what the Factory adds. > > > > > > > > We're talking about getting HBASE-82 into 0.2. Does that interfere > with > > > > > > > this proposal? (I'm thinking that along w/ rows becoming byte arrays > rather > > > Text with a client-supplied Comparator, column qualifiers would shift to > be > > > byte arrays also -- though yeah, implies that if your sort is not > > > byte-lexicographical, yes, the compares can be costly involving two > > > deserializations per compare). > > > > > > > > > > > > > > > > - We catch all writes/deletes and maintain a SortedMap<T, HStoreKey> > > > > > which keeps the column values in order, and maps them back to row > > > > > keys. First cut may just keep all this in memory, but it should be > > > > > backed with MapFile(s). > > > > > > > > > > > > > > > > > > > > > > > Would be sweet if you could leverage the HBase memcache code and > flusher > > > > > > > to do the above. > > > > > > > > > > > This Map would be global for the table? Or per Region? > > > > > > > > A lucene index wouldn't work for you because you want ordering? > > > > > > > > > > > > > > > > > - Add to the hregion the ability to scan through keys in column > order. > > > > > Just iterate through the SortedMap, run a filter on the key, and if > it > > > > > passes do a get on the row. > > > > > > > > > > > > > > > > > > > > > > > You'd be random reading rows. You're OK w/ current performance? (For > > > > > > > sure it will only improve but....). > > > > > > > > > > > > > > > > > > > > - Add a ColumnOrderedClientScanner which will open column order > > > > > scanners to all applicable hregions, and continuously pick row with > > > > > the lowest column value from each of the client scanners. > > > > > > > > > > > > > > > > > > > > > > > This scanner would have a significant client-side component to do the > > > > > > > arbitrage between all regions to figure the lowest column value? If you > had > > > a new type of 'region' -- one denoted by lowest and upper column then > the > > > client-side logic would fade away and your scanner would look like > current > > > scanners. > > > > > > > > > > > > > > > > - Region splits should be easy enough, just a scan through the > > > > > SortedMap to partition. > > > > > > > > > > > > > > > > > > > > > > > Splits would not be row-based and run as they currently do, but rather > > > > > > > sorted-column based? > > > > > > > > > > > > > > > > > > > > Of course, the index could also be used for more efficient querying > on > > > > > the indexed column's values. > > > > > > > > > > Do other users have a need for this functionality? > > > > > > > > > > What do developers think about this? I know hbase is more intended > for > > > > > back-end batch style processing, but we have this need. > > > > > > > > > > > > > > > > > > > > > > > > How are you thinking of adding in this new functionality? Subclassing > > > > > > > HRegionServer? > > > > > > > > > > > St.Ack > > > > > > > > > > > > > > > > > Cheers, > > > > > -clint > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
