Thank you DuyHai.
I was in two minds about large partitions for my app.
I thought upgrading to 3.x would be good and easy option. But now I'm going to work on refactoring my data model :)

2016-10-15 20:38 GMT+09:00 DuyHai Doan <doanduyhai@gmail.com>:
Yes, more or less. The 100Mb is a rule of thumb. No one will blame you for storing 200Mb for example. The figure is just given as an example of order of magnitude

On Sat, Oct 15, 2016 at 1:37 PM, Kant Kodali <kant@peernova.com> wrote:
you mean 100MB (MegaBytes)? Also the data in each of my column is about 1KB so in that case the optimal size 100K columns (since 100K * 1KB = 100MB) right? 

On Sat, Oct 15, 2016 at 4:26 AM, DuyHai Doan <doanduyhai@gmail.com> wrote:
"2) so what is optimal limit in terms of data size?"

--> Usual recommendations for Cassandra 2.1 are:

a. max 100Mb per partition size
b. or up to 10 000 000 physical columns for a partition (including clustering columns etc ...)

Recently, with the work of Robert Stupp (CASSANDRA-11206) and also with the huge enhancement from Michael Kjellman (CASSANDRA-9754) it will be easier to handle huge partition in memory, especially with a reduce memory footprint with regards to the JVM heap.

However, as long as we don't have repair and streaming processes that can be "resumed" in a middle of a partition, the operational pains will still be there. Same for compaction



On Sat, Oct 15, 2016 at 12:00 PM, Kant Kodali <kant@peernova.com> wrote:
1) It will be great if someone can confirm that there is no limit
2) so what is optimal limit in terms of data size?

Finally, Thanks a lot for pointing out all the operational issues!

On Sat, Oct 15, 2016 at 2:39 AM, DuyHai Doan <doanduyhai@gmail.com> wrote:
"But is there still 2B columns limit on the Cassandra code?"

--> I remember some one the committer saying that this 2B columns limitation comes from the Thrift era where you're limited to max  2B columns to be returned to the client for each request. It also applies to the max size of each "page" of data

Since the introduction of the binary protocol and the paging feature, this limitation does not make sense anymore.

By the way, if your partition is too wide, you'll face other operational issues way before reaching the 2B columns limit:

- compaction taking looooong time --> heap pressure --> long GC pauses --> nodes flapping 
- repair & over-streaming, repair session failure in the middle that forces you to re-send the whole big partition --> the receiving node has a bunch of duplicate data --> pressure on compaction
- bootstrapping of new nodes --> failure to stream a partition in the middle will force to re-send the whole partition from the beginning again --> the receiving node has a bunch of duplicate data --> pressure on compaction



On Sat, Oct 15, 2016 at 9:15 AM, Kant Kodali <kant@peernova.com> wrote:
 compacting 10 sstables each of them have a 15GB partition in what duration?

On Fri, Oct 14, 2016 at 11:45 PM, Matope Ono <matope.ono@gmail.com> wrote:
Please forget the part in my sentence.
For more correctly, maybe I should have said like "He could compact 10 sstables each of them have a 15GB partition".
What I wanted to say is we can store much more rows(and columns) in a partition than before 3.6.

2016-10-15 15:34 GMT+09:00 Kant Kodali <kant@peernova.com>:
"Robert said he could treat safely 10 15GB partitions at his presentation" This sounds like there is there is a row limit too not only columns??

If I am reading this correctly 10 15GB partitions  means 10 partitions (like 10 row keys,  thats too small) with each partition of size 15GB. (thats like 15 million columns where each column can have a data of size 1KB).

On Fri, Oct 14, 2016 at 11:30 PM, Kant Kodali <kant@peernova.com> wrote:
"Robert said he could treat safely 10 15GB partitions at his presentation" This sounds like there is there is a row limit too not only columns??

If I am reading this correctly 10 15GB partitions  means 10 partitions (like 10 row keys,  thats too small) with each partition of size 15GB. (thats like 10 million columns where each column can have a data of size 1KB).





On Fri, Oct 14, 2016 at 9:54 PM, Matope Ono <matope.ono@gmail.com> wrote:
Thanks to CASSANDRA-11206, I think we can have much larger partition than before 3.6. 
(Robert said he could treat safely 10 15GB partitions at his presentation. https://www.youtube.com/watch?v=N3mGxgnUiRY)

But is there still 2B columns limit on the Cassandra code?
If so, out of curiosity, I'd like to know where the bottleneck is. Could anyone let me know about it?

Thanks Yasuharu.


2016-10-13 1:11 GMT+09:00 Edward Capriolo <edlinuxguru@gmail.com>:
The "2 billion column limit" press clipping "puffery". This statement seemingly became popular because highly traffic traffic-ed story, in which a tech reporter embellished on a statement to make a splashy article. 

The effect is something like this:
http://www.healthnewsreview.org/2012/08/iced-tea-kidney-stones-and-the-study-that-never-existed/

Iced tea does not cause kidney stones! Cassandra does not store rows with 2 billion columns! It is just not true.






On Wed, Oct 12, 2016 at 4:57 AM, Kant Kodali <kant@peernova.com> wrote:
Well 1) I have not sent it to postgresql mailing lists 2) I thought this is an open ended question as it can involve ideas from everywhere including the Cassandra java driver mailing lists so sorry If that bothered you for some reason.

On Wed, Oct 12, 2016 at 1:41 AM, Dorian Hoxha <dorian.hoxha@gmail.com> wrote:
Also, I'm not sure, but I don't think it's "cool" to write to multiple lists in the same message. (based on postgresql mailing lists rules).
Example I'm not subscribed to those, and now the messages are separated.

On Wed, Oct 12, 2016 at 10:37 AM, Dorian Hoxha <dorian.hoxha@gmail.com> wrote:
There are some issues working on larger partitions.
Hbase doesn't do what you say! You have also to be carefull on hbase not to create large rows! But since they are globally-sorted, you can easily sort between them and create small rows.

In my opinion, cassandra people are wrong, in that they say "globally sorted is the devil!" while all fb/google/etc actually use globally-sorted most of the time! You have to be careful though (just like with random partition)

Can you tell what rowkey1, page1, col(x) actually are ? Maybe there is a way.
The most "recent", means there's a timestamp in there ?

On Wed, Oct 12, 2016 at 9:58 AM, Kant Kodali <kant@peernova.com> wrote:
Hi All,

I understand Cassandra can have a maximum of 2B rows per partition but in practice some people seem to suggest the magic number is 100K. why not create another partition/rowkey automatically (whenever we reach a safe limit that  we consider would be efficient)  with auto increment bigint  as a suffix appended to the new rowkey? so that the driver can return the new rowkey  indicating that there is a new partition and so on...Now I understand this would involve allowing partial row key searches which currently Cassandra wouldn't do (but I believe HBASE does) and thinking about token ranges and potentially many other things..

My current problem is this

I have a row key followed by bunch of columns (this is not time series data)
and these columns can grow to any number so since I have 100K limit (or whatever the number is. say some limit) I want to break the partition into level/pages

rowkey1, page1->col1, col2, col3......
rowkey1, page2->col1, col2, col3......

now say my Cassandra db is populated with data and say my application just got booted up and I want to most recent value of a certain partition but I don't know which page it belongs to since my application just got booted up? how do I solve this in the most efficient that is possible in Cassandra today? I understand I can create MV, other tables that can hold some auxiliary data such as number of pages per partition and so on..but that involves the maintenance cost of that other table which I cannot afford really because I have MV's, secondary indexes for other good reasons. so it would be great if someone can explain the best way possible as of today with Cassandra? By best way I mean is it possible with one request? If Yes, then how? If not, then what is the next best way to solve this? 

Thanks,
kant