Hope I'm not too late here... regarding hot spotting with sequential keys, I'd suggest you read this Sematext blog - http://blog.sematext.com/2012/04/09/hbasewd-avoid-regionserver-hotspotting-despite-writing-records-with-sequential-keys/ They present a nice idea there for this kind of issues.
Good Luck! On Mon, Apr 15, 2013 at 11:18 PM, Ted Yu <yuzhih...@gmail.com> wrote: > bq. write performance would be lower > > The above means poorer performance. > > bq. I could batch them up application side > > Please do that. > > bq. I guess there is no way to turn that off? > > That's right. > > On Mon, Apr 15, 2013 at 11:15 AM, Kireet <kir...@feedly.com> wrote: > > > > > > > > > Thanks for the reply. "write performance would be lower" -> this means > > better? > > > > Also I think I used the wrong terminology regarding batching. I meant to > > ask if it uses the client side write buffer. I would think not since the > > append() method returns a Result. I could batch them up application side > I > > suppose. Append also seems to return the updated value. This seems like a > > lot of unnecessary I/O in my case since I am not immediately interested > in > > the updated value. I guess there is no way to turn that off? > > > > > > On 4/15/13 1:28 PM, Ted Yu wrote: > > > >> I assume you would select HBase 0.94.6.1 (the latest release) for this > >> project. > >> > >> For #1, write performance would be lower if you choose to use Append > (vs. > >> using Put). > >> > >> bq. Can appends be batched by the client or do they execute immediately? > >> This depends on your use case. Take a look at the following method in > >> HTable where you can send a list of actions (Appends): > >> > >> public void batch(final List<?extends Row> actions, final Object[] > >> results) > >> For #2 > >> bq. The other would be to prefix the timestamp row key with a random > >> leading byte. > >> > >> This technique has been used elsewhere and is better than the first one. > >> > >> Cheers > >> > >> On Mon, Apr 15, 2013 at 6:09 AM, Kireet Reddy > <kireet-Teh5dPVPL8nQT0dZR+* > >> *a...@public.gmane.org < > kireet-teh5dpvpl8nqt0dzr%2ba...@public.gmane.org>> > >> wrote: > >> > >> I are planning to create a "scheduled task list" table in our hbase > >>> cluster. Essentially we will define a table with key timestamp and then > >>> the > >>> row contents will be all the tasks that need to be processed within > that > >>> second (or whatever time period). I am trying to do the "reasonably > wide > >>> rows" design mentioned in the hbasecon opentsdb talk. A couple of > >>> questions: > >>> > >>> 1. Should we use append or put to create tasks? Since these rows will > not > >>> live forever, storage space in not a concern, read/write performance is > >>> more important. As concurrency increases I would guess the row lock may > >>> become an issue in append? Can appends be batched by the client or do > >>> they > >>> execute immediately? > >>> > >>> 2. I am a little worried about hotspots. This basic design may cause > >>> issues in terms of the table's performance. Many tasks will execute and > >>> reschedule themselves using the same interval, t + 1 hour for example. > So > >>> many the writes may all go to the same block. Also, we have a lot of > >>> other > >>> data so I am worried it may impact performance of unrelated data if the > >>> region server gets too busy servicing the task list table. I can think > >>> of 2 > >>> strategies to avoid this. One would be to create N different tables and > >>> read/write tasks to them randomly. This may spread load across servers, > >>> but > >>> there is no guarantee hbase will place the tables on different region > >>> servers, correct? The other would be to prefix the timestamp row key > >>> with a > >>> random leading byte. Then when reading from the task list table, > >>> consumers > >>> could scan from any/all possible values of the random byte + current > >>> timestamp to obtain tasks. Both strategies seem like they could spread > >>> out > >>> load, but at the cost of more work/complexity to read tasks from the > >>> table. > >>> Do either of those approaches make sense? > >>> > >>> On the read side, it seems like a similar problem exists in that all > >>> consumers will be reading rows based on the current timestamp. Is this > >>> good > >>> because the block will very likely be cached or bad because the region > >>> server may become overloaded? I have a feeling the answer is going to > be > >>> "it depends". :) > >>> > >>> I did see the previous posts on queues and the tips there - use > zookeeper > >>> for coordination, schedule major compactions, etc. Sorry if these > >>> questions > >>> are basic, I am pretty new to hbase. Thanks! > >>> > >> > >> > > > > >