Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Won't Fix
-
2.2.2
-
None
-
None
Description
Hello
while investigating some OOM errors in Spark 2.2 (here's my call stack), I find the following behavior happening, which I think is weird:
- a request happens to send a partition over network
- this partition is 1.9 GB and is persisted in memory
- this partition is apparently stored in a ByteBufferBlockData, that is made of a ChunkedByteBuffer, which is a list of (lots of) ByteBuffer of 4 MB each.
- the call to toNetty() is supposed to only wrap all the arrays and not allocate any memory
- yet the call stack shows that netty is allocating memory and is trying to consolidate all the chunks into one big 1.9GB array
- this means that at this point the memory footprint is 2x the size of the actual partition (which is huge when the partition is 1.9GB)
Is this transient allocation expected?
After digging, it turns out that the actual copy is due to this method in netty. If my initial buffer is made of more than DEFAULT_MAX_COMPONENTS (16) components it will trigger a re-allocation of all the buffer. This netty issue was fixed in this recent change : https://github.com/netty/netty/commit/9b95b8ee628983e3e4434da93fffb893edff4aa2
As a result, is it possible to benefit from this change somehow in spark 2.2 and above? I don't know how the netty dependencies are handled for spark
NB: it seems this ticket: https://jira.apache.org/jira/browse/SPARK-24307 kinda changed the approach for spark 2.4 by bypassing netty buffer altogether. However as it is written in the ticket, this approach still needs to have the entire block serialized in memory, so this would be a downgrade from fixing the netty issue when your buffer in < 2GB
Thanks!
