Re: Capacity Planning For Large State in YARN Cluster

Posted by Ashish Pokharel on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Capacity-Planning-For-Large-State-in-YARN-Cluster-tp16444p16472.html

Jorn, correct and I suppose that's where we are at this point. RocksDB based backend is definitely looking promising for our use case. Since I haven't gotten a definite no-no on using 30% for YARN cut-off ratio (about 1.8GB from 6GB memory) and off-heap flag turned on, we will continue on that path. Current plan is to increase throughput on input streams - state streams are pretty much processing already and preserved in RocksDB and we can control streams for joining with those states and monitor resource utilizations + join performance. We are seeing 200-500ms processing times with pretty decent amount of logging, which is pretty good for our needs. 

Agree about the way to estimate the size of state and hence one of the reasons of my original question on what others have done. Our states are essentially tuples (few primitive values like string, long and a Map of string and string, which hold about 10-12 keys, values are small - not more than 128 bytes tops). We created a savepoint after processing about 500k records and that's where my estimate came from. I'd be the first one to admit it is not accurate but that's the best we could think of. 

Thanks, Ashish

---

From: Jörn Franke <[hidden email]>
To: Ashish Pokharel <[hidden email]>
Cc: Till Rohrmann <[hidden email]>; user <[hidden email]>
Sent: Sunday, October 29, 2017 6:05 PM
Subject: Re: Capacity Planning For Large State in YARN Cluster

Well you can only performance test it beforehand in different scenarios with different configurations. 

I am not sure what exactly your state holds (eg how many objects etc), but if it is Java objects then 3 times might be a little bit low (depends also how you initially tested state size) - however Flink optimizes this as well. Nevertheless, something like Rocksdb is probably a better solution for larger states.

On 29. Oct 2017, at 21:15, Ashish Pokharel <[hidden email]> wrote:

Hi Till,

I got the same feedback from Robert Metzger over in Stackflow. I have switched my app to use RocksDB and as yes, it did stabilize the app :) 

However, I am still struggling with how to map out my TMs and JMs memory, number of slots per TMs etc. Currently I am using 60 slots with 10 TMs and 60 GB of total cluster memory. Idea was to make the states distributed and approx. 1 GB of memory per slot. I have also changed containerized.heap-cutoff-ratio config to 0.3 to allow for a little room for RocksDB (RocksDB is using basic spinning disk optimized pre-defined configs but we do have SSDs on our Prod machines that we can leverage in future too) and set taskmanager.memory.off-heap to true.It feels more experimental at this point than an exact science :) If there are any further guidelines on how we can plan for this as we open up the flood gates to stream heavy continuous streams, that will be great.

Thanks again,

Ashish

On Oct 27, 2017, at 8:45 AM, Till Rohrmann <[hidden email]> wrote:

Hi Ashish,

what you are describing should be a good use case for Flink and it should be able to run your program.

When you are seeing a GC overhead limit exceeded error, then it means that Flink or your program are creating too many/too large objects filling up the memory in a short time. I would recommend checking your user program to see whether you can avoid unnecessary object instantiations and whether it is possible to reuse created objects.

Concerning Flink's state backends, the memory state backend is currently not able to spill to disk. Also the managed memory is only relevant for DataSet/batch programs and not streaming programs. Therefore, I would recommend you to try out the RocksDB state backend which is able to gracefully spill to disk if the state size should grow too large. Consequently, you don't have to adjust the managed memory settings because they currently don't have an effect on streaming programs. 

My gut feeling is that switching to the RocksDBStateBackend could already solve your problems. If this should not be the case, then please let me know again.

Cheers,

Till

On Fri, Oct 27, 2017 at 5:27 AM, Ashish Pokharel <[hidden email]> wrote:

Hi Everyone,

We have hit a roadblock moving an app at Production scale and was hoping to get some guidance. Application is pretty common use case in stream processing but does require maintaining large number of keyed states. We are processing 2 streams - one of which is a daily burst of stream (normally around 50 mil but could go upto 100 mil in one hour burst) and other is constant stream of around 70-80 mil per hour. We are doing a low level join using CoProcess function between the two keyed streams. CoProcess function needs to refresh (upsert) state from the daily burst stream and decorate constantly streaming data with values from state built using bursty stream. All of the logic is working pretty well in a standalone Dev environment. We are throwing about 500k events of bursty traffic for state and about 2-3 mil of data stream. We have 1 TM with 16GB memory, 1 JM with 8 GB memory and 16 slots (1 per core on the server) on the server. We have been taking savepoints in case we need to restart app for with code changes etc. App does seem to recover from state very well as well. Based on the savepoints, total volume of state in production flow should be around 25-30GB.

At this point, however, we are trying deploy the app at production scale. App also has a flag that can be set at startup time to ignore data stream so we can simply initialize state. So basically we are trying to see if we can initialize the state first and take a savepoint as test. At this point we are using 10 TM with 4 slots and 8GB memory each (idea was to allocate around 3 times estimated state size to start with) but TMs keep getting killed by YARN with a GC Overhead Limit Exceeded error. We have gone through quite a few blogs/docs on Flink Management Memory, off-heap vs heap memory, Disk Spill over, State Backend etc. We did try to tweak managed-memory configs in multiple ways (off/on heap, fraction, network buffers etc) but can’t seem to figure out good way to fine tune the app to avoid issues. Ideally, we would hold state in memory (we do have enough capacity in Production environment for it) for performance reasons and spill over to disk (which I believe Flink should provide out of the box?). It feels like 3x anticipated state volume in cluster memory should have been enough to just initialize state. So instead of just continuing to increase memory (which may or may not help as error is regarding GC overhead) we wanted to get some input from experts on best practices and approach to plan this application better.

Appreciate your input in advance!