[jira] [Commented] (CASSANDRA-5982) OutOfMemoryError when writing text blobs to a very large number of tables

["Jonathan Ellis (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-5982?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13767077#comment-13767077 ] 

Jonathan Ellis commented on CASSANDRA-5982:
-------------------------------------------

This workload presents several challenges to Cassandra:

- MemoryMeter is slow, and we [correctly] limit it to one CFS at a time.  If MemoryMeter is busy measuring a large memtable, we can easily write a lot of data to other CFs (or even the same one in a new memtable).  So we need to pick a more realistic liveRatio default to start with until we can get the first measurements.  (This is not a problem for Ryan's workload here; extremely large blobs have a liveRatio close to 1 anyway.  But it could be a problem for a workload with smaller cells.)
- The MemoryMeter queue is unbounded, and can keep memtables on-heap long after they've been flushed.  We should keep a reference to CFS instead and measure the currently active memtable when the task runs.
- forceFlush acquires Table.switchLock.writeLock temporarily, which means that it will block as long as readLock is held.  In particular, this means it will block for CommitLog.add.  This can actually be worse for PCLE than BCLE, since the former will allow up to 1K unwritten entries on its queue.  If the queue is full of large blobs like the ones in this workload, clearing up space in that queue can take a while.  (The queue itself can also be a significant source of memory consumption!)  We should move forceFlush to a separate executor so it can be truely nonblocking (which will allow MeteredFlusher to impose a moratorium on writes to the other CFs that should be flushed that much faster), and also reduce the CL queue size.  There is no reasonable way to adjust the queue size at runtime, so a configuration setting should be introduced.

IMO, all but the new pre-flush executor changes are reasonable to make in 1.2.10; the last I'd rather keep 2.0-only.
                
> OutOfMemoryError when writing text blobs to a very large number of tables
> -------------------------------------------------------------------------
>
>                 Key: CASSANDRA-5982
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-5982
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>            Reporter: Ryan McGuire
>            Assignee: Jonathan Ellis
>            Priority: Minor
>             Fix For: 1.2.10, 2.0.1
>
>         Attachments: 2000CF_memtable_mem_usage.png, system.log.gz
>
>
> This test goes outside the norm for Cassandra, creating ~2000 column families, and writing large text blobs to them. 
> The process goes like this:
> Bring up a 6 node m2.2xlarge cluster on EC2. This instance type has enough memory (34.2GB) so that Cassandra will allocate a full 8GB heap without tuning cassandra-env.sh. However, this instance type only has a single drive, so data and commitlog are comingled. (This test has also been run m1.xlarge instances which have four drives (but lower memory) and has exhibited similar results when assigning one to commitlog and 3 to datafile_directories.)
> Use the 'memtable_allocator: HeapAllocator' setting from CASSANDRA-5935.
> Create 2000 CFs:
> {code}
> CREATE KEYSPACE cf_stress WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 3}
> CREATE COLUMNFAMILY cf_stress.tbl_00000 (id timeuuid PRIMARY KEY, val1 text, val2 text, val3 text ) ;
> # repeat for tbl_00001, tbl_00002 ... tbl_02000
> {code}
> This process of creating tables takes a long time, about 5 hours, but for anyone wanting to create that many tables, presumably they only need to do this once, so this may be acceptable.
> Write data:
> The test dataset consists of writing 100K, 1M, and 10M documents to these tables:
> {code}
> INSERT INTO {table_name} (id, val1, val2, val3) VALUES (?, ?, ?, ?)
> {code}
> With 5 threads doing these inserts across the cluster, indefinitely, randomly choosing a table number 1-2000, the cluster eventually topples over with 'OutOfMemoryError: Java heap space'.
> A heap dump analysis indicates that it's mostly memtables:
> !2000CF_memtable_mem_usage.png!
> Best current theory is that this is commitlog bound and that the memtables cannot flush fast enough due to locking issues. But I'll let [~jbellis] comment more on that.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira