I would say it's probably due to a significantly larger number of partitions when using the overwrite method - but really you should be seeing similar performance unless one of the schemas ends up generating a lot more disk IO. If you're planning to read the last N values for an event at the same time the widerow schema would be better, otherwise reading N events using the overwrite schema will result in you hitting N partitions. You really need to take into account how you're going to read the data when you design a schema, not only how many writes you can push through.
On 8 June 2016 at 19:02, John Thomas <jthom...@gmail.com> wrote: > We have a use case where we are storing event data for a given system and > only want to retain the last N values. Storing extra values for some time, > as long as it isn’t too long, is fine but never less than N. We can't use > TTLs to delete the data because we can't be sure how frequently events will > arrive and could end up losing everything. Is there any built in mechanism > to accomplish this or a known pattern that we can follow? The events will > be read and written at a pretty high frequency so the solution would have > to be performant and not fragile under stress. > > > > We’ve played with a schema that just has N distinct columns with one value > in each but have found overwrites seem to perform much poorer than wide > rows. The use case we tested only required we store the most recent value: > > > > CREATE TABLE eventyvalue_overwrite( > > system_name text, > > event_name text, > > event_time timestamp, > > event_value blob, > > PRIMARY KEY (system_name,event_name)) > > > > CREATE TABLE eventvalue_widerow ( > > system_name text, > > event_name text, > > event_time timestamp, > > event_value blob, > > PRIMARY KEY ((system_name, event_name), event_time)) > > WITH CLUSTERING ORDER BY (event_time DESC) > > > > We tested it against the DataStax AMI on EC2 with 6 nodes, replication 3, > write consistency 2, and default settings with a write only workload and > got 190K/s for wide row and 150K/s for overwrite. Thinking through the > write path it seems the performance should be pretty similar, with probably > smaller sstables for the overwrite schema, can anyone explain the big > difference? > > > > The wide row solution is more complex in that it requires a separate clean > up thread that will handle deleting the extra values. If that’s the path > we have to follow we’re thinking we’d add a bucket of some sort so that we > can delete an entire partition at a time after copying some values > forward, on the assumption that deleting the whole partition is much better > than deleting some slice of the partition. Is that true? Also, is there > any difference between setting a really short ttl and doing a delete? > > > > I know there are a lot of questions in there but we’ve been going back and > forth on this for a while and I’d really appreciate any help you could give. > > > > Thanks, > > John > -- Kurt Greaves k...@instaclustr.com www.instaclustr.com
