|RE: Cassandra Counters||Roshni Rajagopal||9/24/12 10:27 AM|
I looked at my mail below, and Im rambling a bit, so Ill try to re-state my queries pointwise.
a) what are the performance tradeoffs on reads & writes between creating a standard column family and manually doing the counts by a lookup on a key, versus using counters.
b) whats the current state of counters limitations in the latest version of apache cassandra?
c) with there being a possibilty of counter values getting out of sync, would counters not be recommended where strong consistency is desired. The normal benefits of cassandra's tunable consistency would not be applicable, as re-tries may cause overstating. So the normal use case is high performance, and where consistency is not paramount.
Subject: Cassandra Counters
Date: Mon, 24 Sep 2012 16:21:55 +0530
I'm trying to understand if counters are a good fit for my use case.
Ive watched http://blip.tv/datastax/counters-in-cassandra-5497678 many times over now...
and still need help!
Suppose I have a list of items- to which I can add or delete a set of items at a time, and I want a count of the items, without considering changing the database or additional components like zookeeper,
I have 2 options_ the first is a counter col family, and the second is a standard one
And in the second I can add a new col with every set of items added or deleted. Over time this row may grow wide.
To display the final count, Id need to read the row, slice through all columns and add them.
In both cases the writes should be fast, in fact standard col family should be faster as there's no read, before write. And for CL ONE write the latency should be same.
For reads, the first option is very good, just read one column for a key
For the second, the read involves reading the row, and adding each column value via application code. I dont think there's a way to do math via CQL yet.
There should be not hot spotting, if the key is sharded well. I could even maintain the count derived from the List_Std_CF in a separate column family which is a standard col family with the final number, but I could do that as a separate process immediately after the write to List_Std_CF completes, so that its not blocking. I understand cassandra is faster for writes than reads, but how slow would Reading by row key be...? Is there any number around after how many columns the performance starts deteriorating, or how much worse in performance it would be?
The advantage I see is that I can use the same consistency rules as for the rest of column families. If quorum for reads & writes, then you get strongly consistent values.
In case of counters I see that in case of timeout exceptions because the first replica is down or not responding, there's a chance of the values getting messed up, and re-trying can mess it up further. Its not idempotent like a standard col family design can be.
If it gets messed up, it would need administrator's help (is there a a document on how we could resolve counter values going wrong?)
I believe the rest of the limitations still hold good- has anything changed in recent versions? In my opinion, they are not as major as the consistency question.
-removing a counter & then modifying value - behaviour is undetermined
-special process for counter col family sstable loss( need to remove all files)
-no TTL support
-no secondary indexes
In short, I can recommend counters can be used for analytics or while dealing with data where the exact numbers are not important, or
when its ok to take some time to fix the mismatch, and the performance requirements are most important.
However where the numbers should match , its better to use a std column family and a manual implementation.
Please share your thoughts on this.
|RE: Cassandra Counters||Milind Parikh||9/24/12 11:02 AM|
You would use Cassandra Counters (or other variation of distributed counting) in case of having determined that a centralized version of counting is not going to work.
You'd determine the non_feasibility of centralized counting by figuring the speed at which you need to sustain writes and reads and reconcile that with your hard disk seek times (essentially).
Once you have "proved" that you can't do centralized counting, the second layer of arsenal comes into play; which is distributed counting.
In distributed counting , the CAP theorem comes into life. & in Cassandra, Availability and Network Partitioning trumps over Consistency.
|RE: Cassandra Counters||Roshni Rajagopal||9/24/12 9:53 PM|
Has anyone implemented counting in a standard col family in cassandra, when you can have increments and decrements to the count.
Any comparisons in performance to using counter column families?
|Re: Cassandra Counters||Oleksandr Petrov||9/24/12 10:57 PM|
Maybe I'm missing the point, but counting in a standard column family would be a little overkill.
I assume that "distributed counting" here was more of a map/reduce approach, where Hadoop (+ Cascading, Pig, Hive, Cascalog) would help you a lot. We're doing some more complex counting (e.q. based on sets of rules) like that. Of course, that would perform _way_ slower than counting beforehand. On the other side, you will always have a consistent result for a consistent dataset.
On the other hand, if you use things like AMQP or Storm (sorry to put up my sentence together like that, as tools are mostly either orthogonal or complementary, but I hope you get my point), you could build a topology that makes fault-tolerant writes independently of your original write. Of course, it would still have a consistency tradeoff, mostly because of race conditions and different network latencies etc.
So I would say that building a data model in a distributed system often depends more on your problem than on the common patterns, because everything has a tradeoff.
Want to have an immediate result? Modify your counter while writing the row.
Can sacrifice speed, but have more counting opportunities? Go with offline distributed counting.
Want to have kind of both, dispatch a message and react upon it, having the processing logic and writes decoupled from main application, allowing you to care less about speed.
However, I may have missed the point somewhere (early morning, you know), so I may be wrong in any given statement.
|RE: Cassandra Counters||Roshni Rajagopal||9/24/12 11:36 PM|
Thanks for the reply and sorry for being bull - headed.
Once you're past the stage where you've decided its distributed, and NoSQL and cassandra out of all the NoSQL options,
Now to count something, you can do it in different ways in cassandra.
In all the ways you want to use cassandra's best features of availability, tunable consistency , partition tolerance etc.
Given this, what are the performance tradeoffs of using counters vs a standard column family for counting. Because as I see if the counter number in a counter column family becomes wrong, it will not be 'eventually consistent' - you will need intervention to correct it. So the key aspect is how much faster would be a counter column family, and at what numbers do we start seing a difference.
|Re: Cassandra Counters||Robin Verlangen||9/25/12 12:15 AM|
From my point of view an other problem with using the "standard column family" for counting is transactions. Cassandra lacks of them, so if you're multithreaded updating counters, how will you keep track of that? Yes, I'm aware of software like Zookeeper to do that, however I'm not sure whether that's the best option.
I think you should stick with Cassandra counter column families.
Disclaimer: The information contained in this message and attachments is intended solely for the attention and use of the named addressee and may be confidential. If you are not the intended recipient, you are reminded that the information remains the property of the sender. You must not use, disclose, distribute, copy, print or rely on this e-mail. If you have received this message in error, please contact the sender immediately and irrevocably delete this message and any copies.
2012/9/25 Roshni Rajagopal <roshni_r...@hotmail.com>
|Re: Cassandra Counters||Edward Kibardin||9/25/12 4:22 AM|
I've recently noticed several threads about Cassandra Counters inconsistencies and started seriously think about possible workarounds like store realtime counters in Redis and dump them daily to Cassandra.
So general question, should I rely on Counters if I want 100% accuracy?
|Re: Cassandra Counters||rohit bhatia||9/25/12 5:17 AM|
We use counters in production with Cassandra 1.0.5. Though since our application is sensitive to write latency and we are seeing problems with Frequent Young Garbage Collections, and also we just do increments (decrements have caused problems for some people)
We don't see inconsistencies in our data.
So if you want 99.99% accurate counters, and can manage with eventual consistency. Cassandra works nicely.
|Re: Cassandra Counters||Sylvain Lebresne||9/25/12 6:27 AM|
Even not considering potential bugs, counters being not idempotent, if you get a TimeoutException during a write (which can happen even in relatively normal conditions), you won't know if the increment went in or not (and you have no way to know unless you have an external way to check the value). This is probably fine if you use counters for say real-time analytics, but not if you use 100% accuracy.
|Re: Cassandra Counters||rohit bhatia||9/25/12 6:38 AM|
In a relatively untroubled cluster, even timed out writes go through,
provided no messages are dropped. Which you can monitor on cassandra
nodes. We have 100% consistency on our production servers as we don't
see messages being dropped on our servers.
Though as you mention, there would be no way to "repair" your dropped messages .
|Re: Cassandra Counters||Edward Kibardin||9/25/12 6:42 AM|
@Sylvain and @Rohit: Thanks for your answers.
|Re: Cassandra Counters||Sylvain Lebresne||9/25/12 7:12 AM|
> In a relatively untroubled cluster, even timed out writes go through,This all depends of your definition of "untroubled" cluster, but to be
clear, in a cluster where a node dies (which for Cassandra is not
considered abnormal and will happen to everyone no matter how good
your monitoring is), you have a good change to get TimeoutExceptions
on counter writes while the other nodes of the cluster haven't
detected the failure (which can take a few seconds) AND those writes
won't get through. The fact that Cassandra logs dropped messages or
not has nothing to do with that.
Though I'm happy for you that you achieve 100% consistency, I want to
re-iter that not seeing any log of messages being dropped does not
guarantee that all counter writes did went true: the ones that timeout
may or may have been persisted.