Roshan, Depending on what your cluster setup is and what the resolution of the time stamp is you could do something like this to spread the data around:
<timestamp-LSBs>-<string>-<reverse timestamp> Using the LSBs of the timestamp as a uniform hash, then splitting on all possible hashes would spread things around a bit. If you do this, then all scans must check all hashes for data. On Tue, Nov 27, 2012 at 1:25 PM, Keith Turner <[email protected]> wrote: > > > On Tue, Nov 27, 2012 at 1:22 PM, Roshan Punnoose <[email protected]>wrote: > >> Thanks! >> >> The fact that you are using a binary tree behind the scenes makes perfect >> sense. Btw, what do you use in the standalone (non native) implementation? >> Does it use a TreeMap? >> > > When not using native code, ConcurrentSkipListMap is used. > > >> >> >> On Tue, Nov 27, 2012 at 12:57 PM, Keith Turner <[email protected]> wrote: >> >>> >>> >>> On Tue, Nov 27, 2012 at 12:21 PM, Roshan Punnoose <[email protected]>wrote: >>> >>>> The <string> would most likely be a fixed set of strings that do not >>>> change over time. >>>> >>>> My question is if it is bad to use a reverse index timestamp in the row >>>> id? Will it cause problems with the tablet splitting, compaction, and >>>> performance if the data is always being sent to the top of the tablet? If I >>>> define a split as everything prefixed with <string>, then the ingest will >>>> go to one tablet, but then I add a reverse timestamp in the row, and that >>>> would mean I am always copying data to the top of the tablet. Will this >>>> cause performance issues? Or is it better to append to a tablet? >>>> >>> >>> I do not think it should matter. Inserts go into a C++ STL map on the >>> tablet server if using the nativemap. I think the implementation of that >>> is a balanced binary tree. So I do not think inserting at the beginning vs >>> the end would make difference. That being said, I do not think I have >>> tried this so I do not know if there would be any suprises. I would be >>> interested in hearing about your experiences. >>> >>> >>>> >>>> >>>> On Tue, Nov 27, 2012 at 11:51 AM, Keith Turner <[email protected]>wrote: >>>> >>>>> >>>>> >>>>> Keith >>>>> >>>>> On Tue, Nov 27, 2012 at 10:41 AM, Roshan Punnoose >>>>> <[email protected]>wrote: >>>>> >>>>>> I want to have a table where the row will consist of >>>>>> "<string>-<reverse index timestamp>". But this means that the data is >>>>>> always being prefixed to the beginning of the row (or tablet if the row >>>>>> is >>>>>> large). Will this be a problem for compaction or performance? >>>>> >>>>> >>>>> Can you tell me more about what <string> is? For example is it a hash >>>>> or does it come from the set "foo1","foo2","foo3". How does it change >>>>> over time? I think the answer to your question depends on what <string> >>>>> is. >>>>> >>>>> >>>>>> >>>>>> I don't know if I heard this correctly, but someone once mentioned >>>>>> that making the row id the direct timestamp could cause performance >>>>>> issues >>>>>> because data is always going to one tablet, but also because there is >>>>>> trouble splitting since it always appends to the tablet. Is this true, is >>>>>> it similar to what could happen if I am always prefixing to a tablet? >>>>>> >>>>> >>>>> Yes using a timestamp for a row could cause data from many clients to >>>>> always go to the same tablet, which would be bad for performance on a >>>>> cluster. >>>>> >>>>> >>>>>> >>>>>> Thanks! >>>>>> Roshan >>>>>> >>>>> >>>>> >>>> >>> >> >
