Hello Spark Users, With the recent graduation of Spark to a top level project (grats, btw!), maybe a well timed question. :)
We are at the very beginning of a large scale big data project and after two months of exploration work we'd like to settle on the technologies to use, roll up our sleeves and start to build the system. Spark is one of the forerunners for our technology choice. My question in essence is whether it's a good idea or is Spark too 'experimental' just yet to bet our lives (well, the project's life) on it. The benefits of choosing Spark are numerous and I guess all too obvious for this audience - e.g. we love its powerful abstraction, ease of development and the potential for using a single system for serving and manipulating huge amount of data. This email aims to ask about the risks. I enlist concrete issues we've encountered below, but basically my concern boils down to two philosophical points: I. Is it too much magic? Lots of things "just work right" in Spark and it's extremely convenient and efficient when it indeed works. But should we be worried that customization is hard if the built in behavior is not quite right for us? Are we to expect hard to track down issues originating from the black box behind the magic? II. Is it mature enough? E.g. we've created a pull request<https://github.com/apache/spark/pull/181>which fixes a problem that we were very surprised no one ever stumbled upon before. So that's why I'm asking: is Spark being already used in professional settings? Can one already trust it being reasonably bug free and reliable? I know I'm asking a biased audience, but that's fine, as I want to be convinced. :) So, to the concrete issues. Sorry for the long mail, and let me know if I should break this out into more threads or if there is some other way to have this discussion... 1. Memory management The general direction of these questions is whether it's possible to take RDD caching related memory management more into our own hands as LRU eviction is nice most of the time but can be very suboptimal in some of our use cases. A. Somehow prioritize cached RDDs, E.g. mark some "essential" that one really wants to keep. I'm fine with going down in flames if I mark too much data essential. B. Memory "reflection": can you pragmatically get the memory size of a cached rdd and memory sizes available in total/per executor? If we could do this we could indirectly avoid automatic evictions of things we might really want to keep in memory. C. Evictions caused by RDD partitions on the driver. I had a setup with huge worker memory and smallish memory on the driver JVM. To my surprise, the system started to cache RDD partitions on the driver as well. As the driver ran out of memory I started to see evictions while there were still plenty of space on workers. This resulted in lengthy recomputations. Can this be avoided somehow? D. Broadcasts. Is it possible to get rid of a broadcast manually, without waiting for the LRU eviction taking care of it? Can you tell the size of a broadcast programmatically? 2. Akka lost connections We have quite often experienced lost executors due to akka exceptions - mostly connection lost or similar. It seems to happen when an executor gets extremely busy with some CPU intensive work. Our hypothesis is that akka network threads get starved and the executor fails to respond within timeout limits. Is this plausible? If yes, what can we do with it? In general, these are scary errors in the sense that they come from the very core of the framework and it's hard to link it to something we do in our own code, and thus hard to find a fix. So a question more for the community: how often do you end up scratching your head about cases where spark magic doesn't work perfectly? 3. Recalculation of cached rdds I see the following scenario happening. I load two RDDs A,B from disk, cache them and then do some jobs on them, at the very least a count on each. After these jobs are done I see on the storage panel that 100% of these RDDs are cached in memory. Then I create a third RDD C which is created by multiple joins and maps from A and B, also cache it and start a job on C. When I do this I still see A and B completely cached and also see C slowly getting more and more cached. This is all fine and good, but in the meanwhile I see stages running on the UI that point to code which is used to load A and B. How is this possible? Am I misunderstanding how cached RDDs should behave? And again the general question - how can one debug such issues? 4. Shuffle on disk Is it true - I couldn't find it in official docs, but did see this mentioned in various threads - that shuffle _always_ hits disk? (Disregarding OS caches.) Why is this the case? Are you planning to add a function to do shuffle in memory or are there some intrinsic reasons for this to be impossible? Sorry again for the giant mail, and thanks for any insights! Andras
