> We’ve actually had several customers where we’ve done the opposite - split large clusters apart to separate uses cases
We do something similar but for a single application. We're functionally sharding data to different clusters from a single application. We can have different server classes for different types of workloads, we can grow and size clusters accordingly, and we also do things like time sharding so that we can let at-rest data go to cheaper storage options. I agree with the general sentiment here that (at least as it stands today) a monolithic cluster for many applications does not compete to per-application clusters unless cost is no issue. At our scale, the terabytes of C* data we take in per day means that even very small cost savings really add up at scale. And even where cost is no issue, the additional isolation and workload tailoring is still highly valuable. On Wed, Feb 22, 2017 at 12:01 PM Edward Capriolo <edlinuxg...@gmail.com> wrote: > > > On Wed, Feb 22, 2017 at 1:20 PM, Abhishek Verma <ve...@uber.com> wrote: > > We have lots of dedicated Cassandra clusters for large use cases, but we > have a long tail of (~100) of internal customers who want to store < 200GB > of data with < 5k qps and non-critical data. It does not make sense to > create a 3 node dedicated cluster for each of these small use cases. So we > have a shared cluster into which we onboard these users. > > But once in a while, one of the customers will run a ingest job from HDFS > which will pound the shared cluster and break our SLA for the cluster for > all the other customers. Currently, I don't see anyway to signal back > pressure to the ingestion jobs or throttle their requests. Another example > is one customer doing a large number of range queries which has the same > effect. > > A simple way to avoid this is to throttle the read or write requests based > on some quota limits for each keyspace or user. > > Please see replies inlined: > > On Mon, Feb 20, 2017 at 11:46 PM, vincent gromakowski < > vincent.gromakow...@gmail.com> wrote: > > Aren't you using mesos Cassandra framework to manage your multiple > clusters ? (Seen a presentation in cass summit) > > Yes we are using https://github.com/mesosphere/dcos-cassandra-service and > contribute heavily to it. I am aware of the presentation ( > https://www.youtube.com/watch?v=4Ap-1VT2ChU) at the Cassandra summit as I > was the one who gave it :) > This has helped us automate the creation and management of these clusters. > > What's wrong with your current mesos approach ? > > Hardware efficiency: Spinning up dedicated clusters for each use case > wastes a lot of hardware resources. One of the approaches we have taken is > spinning up multiple Cassandra nodes belonging to different clusters on the > same physical machine. However, we still have overhead of managing these > separate multi-tenant clusters. > > I am also thinking it's better to split a large cluster into smallers > except if you also manage client layer that query cass and you can put some > backpressure or rate limit in it. > > We have an internal storage API layer that some of the clients use, but > there are many customers who use the vanilla DataStax Java or Python > driver. Implementing throttling in each of those clients does not seem like > a viable approach. > > Le 21 févr. 2017 2:46 AM, "Edward Capriolo" <edlinuxg...@gmail.com> a > écrit : > > Older versions had a request scheduler api. > > I am not aware of the history behind it. Can you please point me to the > JIRA tickets and/or why it was removed? > > On Monday, February 20, 2017, Ben Slater <ben.sla...@instaclustr.com> > wrote: > > We’ve actually had several customers where we’ve done the opposite - split > large clusters apart to separate uses cases. We found that this allowed us > to better align hardware with use case requirements (for example using AWS > c3.2xlarge for very hot data at low latency, m4.xlarge for more general > purpose data) we can also tune JVM settings, etc to meet those uses cases. > > There have been several instances where we have moved customers out of the > shared cluster to their own dedicated clusters because they outgrew our > limitations. But I don't think it makes sense to move all the small use > cases into their separate clusters. > > On Mon, 20 Feb 2017 at 22:21 Oleksandr Shulgin < > oleksandr.shul...@zalando.de> wrote: > > On Sat, Feb 18, 2017 at 3:12 AM, Abhishek Verma <ve...@uber.com> wrote: > > Cassandra is being used on a large scale at Uber. We usually create > dedicated clusters for each of our internal use cases, however that is > difficult to scale and manage. > > We are investigating the approach of using a single shared cluster with > 100s of nodes and handle 10s to 100s of different use cases for different > products in the same cluster. We can define different keyspaces for each of > them, but that does not help in case of noisy neighbors. > > Does anybody in the community have similar large shared clusters and/or > face noisy neighbor issues? > > > Hi, > > We've never tried this approach and given my limited experience I would > find this a terrible idea from the perspective of maintenance (remember the > old saying about basket and eggs?) > > What if you have a limited number of baskets and several eggs which are > not critical if they break rarely. > > > What potential benefits do you see? > > The main benefit of sharing a single cluster among several small use cases > is increasing the hardware efficiency and decreasing the management > overhead of a large number of clusters. > > Thanks everyone for your replies and questions. > > -Abhishek. > > > I agree with these assertions. On one hand I think about a "managed > service" like say Amazon DynamoDB. They likely start with very/very/very > large footprints. IE they commission huge clusters of the fastest SSD > hardware. Next every application/user has a quota. They always can control > the basic load because they control the quota. > > Control on the hardware level makes sense, but then your unit of > management is "a cluster" . Users do not have a unified API anymore, they > have switch statements, this data in cluster x, this data cluster y. You > still end up in cases where degenerate usage patterns affect others. > > With Cassandra it would be nice if these controls were build into the API. > This could also help you build your own charge back model in the > enterprise. Sure as someone pointed out rejecting reads stinks for that > user. But then again someone has to decide who and how pays for the > hardware. > > For example, imagine a company with 19 business units all using the same > Cassandra cluster. One business unit might account for 90% of the storage, > but 1% of the requests. Another business unit might be 95% of the requests, > but 1% the data. How do you come up with a billing model? For the customer > with 95% of the requests their "cost" on the systems is young generation > GC, network. > > Datastax enterprise had/has a concept of "the analytic dc". The concept > is "real time goes here" and "analytic goes there" with the right resource > controls you could get much more fine grained then that. It will never be > perfect there will always be that random abuser with the "aggregate allow > filtering query" but there are ways to move in a more managed direction. >
