Hi, We are internally exploring adding support for dynamically changing the resource profile of a stage based on runtime characteristics. This includes failures due to OOM and the like, slowness due to excessive GC, resource wastage due to excessive overprovisioning, etc. Essentially handles scale up and scale down of resources. Instead of baking these into the scheduler directly (which is already complex), we are modeling it as a plugin - so that the 'business logic' of how to handle task events and mutate state is pluggable.
The main limitation I find with mutating only the cores is the limits it places on what kind of problems can be solved with it - and mutating resource profiles is a much more natural way to handle this (spark.task.cpus predates RP). Regards, Mridul On Wed, Jan 17, 2024 at 9:18 AM Tom Graves <tgraves...@yahoo.com.invalid> wrote: > It is interesting. I think there are definitely some discussion points > around this. reliability vs performance is always a trade off and its > great it doesn't fail but if it doesn't meet someone's SLA now that could > be as bad if its hard to figure out why. I think if something like this > kicks in, it needs to be very obvious to the user so they can see that it > occurred. Do you have something in place on UI or something that indicates > this? The nice thing is also you aren't wasting memory by increasing it for > all tasks when maybe you only need it for one or two. The downside is you > are only finding out after failure. > > I do also worry a little bit that in your blog post, the error you pointed > out isn't a java OOM but an off heap memory issue (overhead + heap usage). > You don't really address heap memory vs off heap in that article. Only > thing I see mentioned is spark.executor.memory which is heap memory. > Obviously adjusting to only run one task is going to give that task more > overall memory but the reasons its running out in the first place could be > different. If it was on heap memory for instance with more tasks I would > expect to see more GC and not executor OOM. If you are getting executor > OOM you are likely using more off heap memory/stack space, etc then you > allocated. Ultimately it would be nice to know why that is happening and > see if we can address it to not fail in the first place. That could be > extremely difficult though, especially if using software outside Spark that > is using that memory. > > As Holden said, we need to make sure this would play nice with the > resource profiles, or potentially if we can use the resource profile > functionality. Theoretically you could extend this to try to get new > executor if using dynamic allocation for instance. > > I agree doing a SPIP would be a good place to start to have more > discussions. > > Tom > > On Wednesday, January 17, 2024 at 12:47:51 AM CST, kalyan < > justfors...@gmail.com> wrote: > > > Hello All, > > At Uber, we had recently, done some work on improving the reliability of > spark applications in scenarios of fatter executors going out of memory and > leading to application failure. Fatter executors are those that have more > than 1 task running on it at a given time concurrently. This has > significantly improved the reliability of many spark applications for us at > Uber. We made a blog about this recently. Link: > https://www.uber.com/en-US/blog/dynamic-executor-core-resizing-in-spark/ > > At a high level, we have done the below changes: > > 1. When a Task fails with the OOM of an executor, we update the core > requirements of the task to max executor cores. > 2. When the task is picked for rescheduling, the new attempt of the > task happens to be on an executor where no other task can run concurrently. > All cores get allocated to this task itself. > 3. This way we ensure that the configured memory is completely at the > disposal of a single task. Thus eliminating contention of memory. > > The best part of this solution is that it's reactive. It kicks in only > when the executors fail with the OOM exception. > > We understand that the problem statement is very common and we expect our > solution to be effective in many cases. > > There could be more cases that can be covered. Executor failing with OOM > is like a hard signal. The framework(making the driver aware of > what's happening with the executor) can be extended to handle scenarios of > other forms of memory pressure like excessive spilling to disk, etc. > > While we had developed this on Spark 2.4.3 in-house, we would like to > collaborate and contribute this work to the latest versions of Spark. > > What is the best way forward here? Will an SPIP proposal to detail the > changes help? > > Regards, > Kalyan. > Uber India. >
