Note that Spark 1.0.0 makes it possible to trivially submit spark jobs to an existing Hadoop cluster.
It leverages HDFS to distribute archives (e.g. your app JAR) and store results / state / logs, and YARN to schedule itself and acquire compute resources.
It's pretty amazing to see how you use Spark's API to write functional applications that are then distributed across multiple executors (e.g. when you use Spark's "filter" or a "map" operations, then the operation potentially gets distributed and distributed on totally different nodes).
I use spark a lot and my experience has been quite the opposite. The queries I run against spark are billions of events and results are sub-second.
I could only speculate as to what this users issues were. One difference between hadoop and spark is that it is more sensitive in that you sometimes need to tell it how many tasks to use. In practice it is no big deal at all.
Perhaps the user was running into this- the data for a task in spark runs all in memory, whereas hadoop will load and spill to disk within a task. So if you give a single hadoop reducer 1TB of data, it will complete after a very long time. In spark if you did this you would need to have 1TB of memory on the executor. I wouldn't give an executor/JVM anything over 10GB. So if you have lots of memory, just be sure to balance it with cores and executors.
I have seen spark use up all the inodes on systems before. A job with 1000 map and 1000 reduce tasks would create 1M spill files on disk. However that was on an earlier version of spark and I was using ext3. I think this has since been improved.
>The queries I run against spark are billions of events and results are sub-second.
This is interesting, I haven't gotten Spark to do anything at all in less than a second. How big is this dataset (what does each event consist of)? How is the data stored? How many machines / cores are running across? What sort of queries are you running?
>I could only speculate as to what this users issues were.
I'm the author of the above post and unfortunately I can also "only speculate" what my issues were. Maybe Spark doesn't like 100x growth in the size of an RDD using flatMap? Maybe large-scale joins don't work well? Who knows. The problem, however, definitely doesn't seem to be anything from the tuning guide(s).
> How big is this dataset (what does each event consist of)?
Standard clickstream data, maybe 50-ish parameters per event.
> What sort of queries are you running?
> How is the data stored?
Depends on the use-case. For sub-second adhoc queries we go against bitmap indexes. Other queries we uses RDD.cache() after a group/cogroup and answer queries directly from that. For other queries we go hit ORC files. Spark is very memory sensitive compared to hadoop, so using a columnar store and only pulling out the data that you absolutely need goes a very long way. Minimizing cross-communication and shuffling is key to achieving sub-second. It's impossible to achieve that if you're waiting for TB of data to shuffle around =)
> How many machines / cores are running across?
Depends on the use case. Clusters are 10-30 machines, some we run virtual on open stack. We will grow our 30 node cluster in 6mo.
> Maybe Spark doesn't like 100x growth in the size of an RDD using flatMap
You may actually just need to proportionally scale the number of partitions for that particular task by the same amount. Also when possible use mapPartitions, it is very memory efficient compared to map/flatMap.
> Maybe large-scale joins don't work well
Keep in mind that what ever happens per task happens all in memory. For large joins I created a "bloom join" implementation (not currently open source =( ) that does this efficiently. It takes two passes at the data, but minimizes what is shuffled.
> Maybe Spark doesn't like 100x growth in the size of an RDD using flatMap?
I'd be interested to hear more about your use case and the problems you encountered. It's possible that you need to do some kind of .coalesce() operation to rebalance the partitions if you have unbalanced partition sizes.
Well, the RDD is initially partitioned using a RangePartitioner over a dense key space of Longs. Each element is then expanded ~100x (each object is significantly smaller than the original value). So the total memory footprint and skew of the expanded RDD shouldn't, theoretically, be a problem.
My experience using spark has also been nothing but positive. I recently built a similarity-based recommendation on Spark (https://github.com/evancasey/sparkler), and found it to be significantly faster than comparable implementations on Hadoop.
subprotocol's point about specifying the number of tasks/data partitions to use is true - you need to manually set this in order to get good results even on a small dataset. However, other than that, spark will give you good results pretty much out of the box. More advanced features such as broadcast objects, cache operations, and custom serializers will further optimize your application, but are not critical when first starting out as the author seems to believe.
I'm really curious to find out in what situations Spark actually works for people. So far, no one in my lab seems to be having a terribly productive time using it. Maybe it's better for simple numerical computations? How large are the datasets you're working with?
I did most of my benchmarking with the 10M MovieLens Dataset http://grouplens.org/datasets/movielens/ consisting of 10 million movie ratings on 10,000 movies from 72,000 users. So not necessarily "big data", but big enough to warrant a distributed approach.
Spark is ideally suited for iterative, multi-stage jobs. In theory, anything that requires doing multiple operations an a working dataset (i.e. graph processing, recommender systems, gradient descent) will do well on Spark due to the in-memory data caching model. This post explains some of the applications Spark is well-suited for: http://www.quora.com/Apache-Spark/What-are-use-cases-for-spa...
So the central piece of data is something like a 10 million element RDD of (UserId, (MovieId, Rating))? If so, it sounds like that data would fit into a single in-memory sparse array, how does Spark's performance compare with a local implementation?
By comparison, I'm trying (and failing) to work with RDDs of 100+ billion elements.
Storm is generally more of a dataflow "per event" real/near time computation system (with each event flowing through N spouts and bolts) whereas Spark is more of an in-memory data processing system (with Spark streaming being the "equivalent" to the storm system).
Spark's abstractions are indeed really nice; a Spark job is much more readable than the same thing expressed in raw MapReduce, as that post acknowledges at the end.
I can't really comment on or rebut "my code runs slow and I don't know why", except to say that Spark performance has been great when I've used it. But yeah, if the abstraction should fail (and again all I can say is it hasn't for me) then I can imagine it's not much fun to debug performance and there's no distributed profiler (though I think you'd be in much the same boat with vanilla Hadoop).
Spark is an interesting technology, from what I've heard it doesn't actually have traction in industry yet though.
Anyone here actually using it in production? I know it's blazing fast etc, and I like it as a map reduce replacement. It has all the makings of a great distributed system, I'm still waiting to see a major deployment yet..
I don't know what you would count as major deployment, but I've deployed a 30-node cluster on HW for running sub-second real-time adhoc queries. I've also run many smaller 10-20 node virtual clusters on open stack. It is a rock solid platform. Our hosted ops loves it because it just works.
The amazing thing about spark is how insanely expressive and hackable it is. The best way I can describe it is this:
* Hadoop: You spend all of your time telling it how to do what you want (it is the assembly language of bigdata)
* Spark: you spend your time telling it what you want, and it just does it
> It is a rock solid platform. Our hosted ops loves it because it just works.
I'm really confused by how different our experiences have been. Above you said you wrote your own shuffle implementation, presumably that was prompted by poor performance at some point. And, when you encountered that poor performance, you presumably also saw what happens to Spark when it's overwhelmed: a sea of exceptions. In a short period of time I've encountered lots of the following:
- FileNotFound exceptions when shuffle files couldn't be created
- Too many open file handles (also related to shuffle files)
- Infinite procession of out-of-memory errors on a cluster with 12TB of memory.
- Executor disconnected
- Weird akka errors
- Mysterious serialization errors (Map.values isn't serializable, making a nested partitioner class for some reason didn't work)
These errors are sometimes recoverable and other times kill all the workers on the cluster. Did none of these things happen to your team?
This does help actually. And yes: it doesn't have to be a 1000 node cluster or anything crazy. I've just talked to a lot of people at bigger companies and they've all said it falls over yet.
Thanks for posting this. I'm starting to get a feel for when Spark is usable-- you need an underlying indexed data store which lets you fetch small subsets of your data into RDDs (or, your data can be tiny to begin with). We've been trying to use Spark on input sizes which, while smaller than our cluster's available memory, are probably too big for Spark to handle (> 1TB).
These guys look to be doing some nice work integrating Cassandra and Spark http://blog.tuplejump.com/
They've piggybacked on the Cassandra clustering using a java agent to run the Spark masters. Doesn't seem to be a realease available yet though.
I'm curious about this too--clj-spark didn't work for me so I'm currently prototyping a job using the Java bindings in Clojure. If I end up wrapping the Java bindings in a useful way I would consider putting together some kind of release if there's community interest in that.
However, shouldn't there be a much more flexible approach where you just send your functions to an execution server (just like an agent)? You might want to define some keywords to refer to previously used functions or data.
Then again, such an agent would be pretty much a REPL, so you might just want to ssh to a REPL that does load balancing and has sub-repls (on other machines) that fail over.
For new entrants, here's an email I sent out to some colleagues of mine just getting into ML. I'm wrapping up a project that's using Mahout, and am getting into Spark & MLlib now. I've regurgitated this on reddit already.
I've been following Apache Spark [0], a new-ish Apache project created by UC Berkeley to replace Hadoop MapReduce [1], for about a month now; and, I finally got around to spending some time with it last night and earllllllly this morning.
Added into the Spark mix about a year ago was a strong Machine Learning library (MLlib) [2] similar to Mahout [3] that promises much better performance (comparable/better than Matlab [4]/Vowpal Wabbit [5])
MLlib is a lower level library, which offers a lot of control/power for developers. However, Berkeley's Amplab has also created a higher level abstraction layer for end users called MLI [6]. It's still being actively developed, and although updates are in the works, they haven't been made available to the public repository for a while [7]
Getting up to speed with Spark itself is really pain-free compared to some tools like Mahout etc. There's a quick-start guide for Scala [8], a getting started guide for Spark [9], and lots of other learning/community resources available for Spark [10] [11]
It leverages HDFS to distribute archives (e.g. your app JAR) and store results / state / logs, and YARN to schedule itself and acquire compute resources.
It's pretty amazing to see how you use Spark's API to write functional applications that are then distributed across multiple executors (e.g. when you use Spark's "filter" or a "map" operations, then the operation potentially gets distributed and distributed on totally different nodes).
Great tool — exciting to see it reach 1.0.0!